{ "converter_version": 17, "metadata_schema_version": 3, "paddy_version": "0.2.0.dev0", "topic_map_version": 2, "sides": [ "left", "right" ], "image_gap_policy": "hold", "image_tolerance_s": 0.016666666666666666, "hold_max_age_s": 0.2, "action_hold_s": 0.2, "rig_hash": "38d66c80622e7deb", "cameras_path": "meta/cameras.json", "cameras": { "camera_bottom": { "source_encoding": "jpeg" }, "camera_left": { "source_encoding": "jpeg" }, "camera_right": { "source_encoding": "jpeg" }, "camera_top": { "source_encoding": "jpeg" } }, "episode_sidecars": "meta/paddy/episodes/{trial_key}.json", "frame_age": "paddy.frame_age.{camera}", "action_valid": "paddy.action_valid", "action_valid_sources": [ "target_age", "safety_state_armed", "rejected_count" ], "action_space": "joints", "target_pose": "paddy.target_pose.{side}", "gripper_units": "normalized", "gripper_scale": "0 closed, 1 open", "quaternions": "sign-continuous per episode, first qw >= 0", "engaged": "paddy.engaged", "scaling": "paddy.scaling", "engagement_hold_s": 0.2, "moving": "paddy.moving", "motion": { "speed_window_s": 0.2, "idle_tcp_speed_m_s": 0.005, "idle_joint_speed_rad_s": 0.02, "idle_gripper_change": 0.02, "keep_min_s": 1.0, "keep_gap_s": 0.5 }, "keep_ranges": "meta/paddy/keep_ranges.json", "control_source": "paddy.control_source", "control_source_values": { "0": "nobody", "1": "human", "2": "policy" }, "clock": "ROS header.stamp: the recorder host's system clock as UTC epoch seconds (POSIX); every stream, RealSense frames included (driver global_time), is on it", "utc_s": "paddy.utc_s", "t_start_utc_s": "paddy/t_start_utc_s", "video_rate_hz": 30, "control_rate_source": "_arm/target_pose header stamps, median interval per episode, median over the episodes", "grippers": { "left": { "model": "g2", "slot": "left_gripper", "unit": "fraction, 0 closed .. 1 open", "opening_m": [ 0.0, 0.084 ] }, "right": { "model": "g2", "slot": "right_gripper", "unit": "fraction, 0 closed .. 1 open", "opening_m": [ 0.0, 0.084 ] } }, "kinematics_hash": "8ddf69d95926fa4ff3b4e29ec78743830d33af903f417b1ffe76b7a02349de5a", "robot_description": "meta/paddy/robot_description.urdf", "episode_uuid": "paddy/episode_uuid", "station_id": "paddy/station_id", "action_issued": "paddy.action_issued", "command_sent": "paddy.command_sent.{side}", "subtask_index": "paddy.subtask_index", "rating": "paddy/rating", "tasks": { "Pick cable from one bin and sort into another bin": { "paraphrases": [] } }, "modality": "meta/modality.json", "stats": { "quantiles": [ "q01", "q10", "q50", "q90", "q99" ], "method": "power-of-two histogram", "bins": 2048, "images": "exact (256-level histogram per channel)" }, "options": { "fps": 30, "crf": 0, "image_tolerance_s": null, "hold_max_age_s": 0.2, "action_hold_s": 0.2, "image_gap": "hold", "cameras": null, "sides": null, "task": null, "robot_type": null, "topic_map_version": null, "overwrite": true, "depth": true, "raw_depth": true, "depth_format": "both", "gop": 10, "gripper_units": "normalized", "rotvec": true, "rot6d": true, "joint_velocity": true, "joint_effort": true, "camera_pose": true, "camera_pose_time": "row", "color_source": "auto", "action_space": "joints", "pose_lead": "auto", "tcp_pose": "urdf", "motion": { "speed_window_s": 0.2, "idle_tcp_speed_m_s": 0.005, "idle_joint_speed_rad_s": 0.02, "idle_gripper_change": 0.02, "keep_min_s": 1.0, "keep_gap_s": 0.5 }, "stereo": true }, "depth": { "format": "both", "path": "depth/{video_key}/chunk-{chunk_index:03d}/file-{file_index:03d}.mkv", "keys": { "camera_bottom": "observation.depth.camera_bottom", "camera_top": "observation.depth.camera_top", "camera_left": "observation.depth.camera_left", "camera_right": "observation.depth.camera_right" }, "codec": "ffv1", "pix_fmt": "gray16le", "container": "matroska", "units_m": 0.001, "cameras": [ "camera_bottom", "camera_top", "camera_left", "camera_right" ], "fill_policy": "zero", "timeline": "color", "lerobot": { "key": "observation.images.{camera}_depth", "keys": { "camera_bottom": "observation.images.camera_bottom_depth", "camera_top": "observation.images.camera_top_depth", "camera_left": "observation.images.camera_left_depth", "camera_right": "observation.images.camera_right_depth" }, "codec": "hevc", "pix_fmt": "gray12le", "container": "mp4", "depth_min": 0.01, "depth_max": 10.0, "shift": 3.5, "use_log": true, "unit": "mm", "lossless_in_depth": false, "code_step_mm": { "0.3": 1.25, "0.6": 1.35, "1.0": 1.48 }, "no_data": "code 0 (0 mm is below depth_min); LeRobot loads it as depth_min (10 mm)", "precision_note": "LeRobot's 12-bit log code (DepthEncoderConfig defaults), HEVC lossless on the codes: one code is 1.25 mm at 0.3 m, 1.35 mm at 0.6 m and 1.48 mm at 1.0 m, so a loaded value is within half a code plus LeRobot's rounding to whole mm of the recorded one; depth beyond 10 m loads as 10 m. The FFV1 file (format ffv1 or both) keeps the exact millimetres." }, "raw": { "keys": { "camera_bottom": "observation.depth_raw.camera_bottom", "camera_top": "observation.depth_raw.camera_top", "camera_left": "observation.depth_raw.camera_left", "camera_right": "observation.depth_raw.camera_right" }, "cameras": [ "camera_bottom", "camera_top", "camera_left", "camera_right" ], "grid": "depth", "calibration": "meta/cameras.json .depth, color_T_depth" } }, "stereo": { "key": "observation.images.{camera}_stereo_{side}", "cameras": { "camera_top": { "left": "observation.images.camera_top_stereo_left", "right": "observation.images.camera_top_stereo_right" } }, "streams": { "left": "infra1", "right": "infra2" }, "pairing": "same header stamp", "pair_stamp_tolerance_s": 0.001, "timeline": "color", "frame_age": "paddy.frame_age.{camera}_stereo_{side}", "channels": 1, "encoding": "the gray image as three equal channels, H.264 yuv420p like color, at the pair's own size", "rotation": "never turned (image_rotation_deg applies to color and depth only)", "calibration": "meta/cameras.json cameras..stereo", "calibration_by_camera": { "camera_top": { "source": "measured", "file": "meta/paddy/stereo_calibration_camera_top.json", "rms_px": 0.4323, "held_out_sampson_p95_px": 0.1683, "calibrated_at": "2026-11-17T16:20:58+00:00" } } }, "control_rate_hz": { "left": 67.719, "right": 68.573 }, "conventions": { "version": 1, "action": { "feature": "action", "space": "joints", "meaning": "per arm the joint command, zero-order held, then the gripper command: the quantity a policy that drives the arms through target_joints outputs. The TCP target the teleop controller published is paddy.target_pose.; the pose the follower sent is paddy.command_sent.", "joints": "absolute positions of the URDF joints joint1..jointN, base to wrist, in observation.state's order; not a delta, not a TCP pose", "joint_unit": "rad", "joints_source": { "left": "command_sent", "right": "command_sent" }, "joints_sources": { "command_sent": "the joint angles the follower sent (command_sent joints_rad: a set_servo_angle call's target or, in servo mode, a set_servo_angle_j setpoint; the xArm7 joint path, GELLO and policy arms), held with action_hold_s; timing.command_methods says which", "target_joints": "a leader arm's recorded joint targets (target_joints, GELLO), held with action_hold_s", "next_state": "the next row's measured joints (the last row its own): the usual stand-in for a position-controlled arm's joint command when none was recorded (a mode-7 arm, older bags)" }, "joints_source_per_episode": "conversion.action_source. of meta/paddy/episodes/.json", "before_first_command": "the row's measured joints (observation.state), paddy.action_valid false: the arm stays where it is until its first command, as the gripper command holds the measured opening", "layout_per_arm": { "left": [ "joint1", "joint2", "joint3", "joint4", "joint5", "joint6", "joint7", "gripper_cmd" ], "right": [ "joint1", "joint2", "joint3", "joint4", "joint5", "joint6", "joint7", "gripper_cmd" ] }, "frames": { "left": "Left_link_base", "right": "Right_link_base" }, "tcp_frames": { "left": "Left_link_tcp", "right": "Right_link_tcp" }, "arm_order": [ "left", "right" ], "gripper": { "unit": "fraction, 0 closed .. 1 open", "physical": { "left": { "model": "g2", "source": "metadata", "state_closed": 0.84, "state_open": 0.0, "cmd_closed": -10.0, "cmd_open": 850.0 }, "right": { "model": "g2", "source": "metadata", "state_closed": 0.84, "state_open": 0.0, "cmd_closed": -10.0, "cmd_open": 850.0 } } }, "already_scaled": true, "scaling_note": "action is in robot units after the controller's motion scaling and rate limit; paddy.scaling is the human-input gain in force (informational, not a policy input)" }, "timing": { "row_time": "t_k = t0 + k / fps (t0 = the latest first stamp over the arms' joint states and the cameras' color frames; paused spans removed); timestamp = k / fps; paddy.utc_s and the sidecar's conversion.t_start_s give the absolute time", "rule": "row k holds the last joint command stamped at or before t_k (zero-order hold, no shift): the command issued at t_k; with joints_source next_state the measured joints of row k+1, the position the arm reached next", "stamp_meaning": "command_sent: header.stamp of the command_sent message, when the follower's motion call (set_servo_angle, or set_servo_angle_j in servo mode) returned; target_joints: header.stamp of the leader's target; next_state: the next row's slot time", "execution": "the follower validates each target and commands the arm from it; how depends on the arm's control mode, named per command by command_sent.method (per arm command_methods, per episode the sidecar's conversion.streams.command_sent..methods) and described in execution_by_method. The issue time of each command is header.stamp of the command_sent stream", "command_methods": { "left": [ "set_servo_angle_j" ], "right": [ "set_servo_angle_j" ] }, "execution_by_method": { "set_servo_angle_j": "servo mode (xArm mode 1): the follower takes each accepted target as joints (its own solution of a TCP target, or a leader's joint target), steps a joint setpoint toward the newest one on a fixed tick (rates.control.servo_rate_hz), each joint inside the stream's speed and acceleration caps, and sends every setpoint with set_servo_angle_j; the control box executes the last setpoint it received and plans nothing. command_sent has one entry per setpoint: joints_rad is the setpoint, the pose and target_stamp are those of the target it steps toward" }, "action_hold_s": 0.2, "valid_mask": "paddy.action_valid", "issued_mask": "paddy.action_issued" }, "rates": { "video_fps": 30, "control": { "teleop_target_hz_nominal": "50 with the target buffer on, else the Quest sample rate (about 50 to 72)", "target_buffer_ms": { "left": 0.0, "right": 0.0 }, "servo_rate_hz": { "left": 200.0, "right": 200.0 }, "measured_median_hz": { "action_joints.left": 199.16, "action_joints.right": 225.579, "command_sent.left": 199.16, "command_sent.right": 225.579, "target.left": 67.719, "target.right": 68.573 }, "measured_per_episode": "sidecar conversion.streams.target..rate_hz and conversion.streams.command_sent..rate_hz" }, "observation_state_hz": 100, "eef_pose_hz": 100, "safety_state_hz": 10 }, "observation_state": { "joint_unit": "rad", "joint_order": "per arm the URDF joints joint1..jointN, base to wrist, then the gripper fraction; arms in arm_order", "gripper_unit": "fraction, 0 closed .. 1 open", "gripper_opening_m": "per arm, meta/paddy/info.json grippers..opening_m: the opening between the jaws at fraction 0 and at fraction 1, metres; the fraction is linear in it" }, "eef_pose": { "feature": "observation.eef_pose.", "pose": "measured TCP pose in the arm base frame, m and quaternion xyzw, zero-order held", "mode": "urdf", "source": { "left": "urdf_fk", "right": "urdf_fk" }, "sources": "urdf_fk: the recorded URDF's base_T_tcp at the recorded joint states (100 Hz), the kinematics the wrist cameras' paddy.camera_pose and the teleop targets use; reported: the pose the arm's control box reports (about 50 Hz, its own kinematics, which can differ from the URDF's by several mm: sidecar conversion.kinematics..tcp_vs_urdf_mm); mixed: episodes differ, see each sidecar's conversion.tcp_pose.", "stamps": "urdf_fk: the joint states' stamps. reported: the control box reports the pose ahead of the arm; each sample is restamped by that lead, measured per episode against the URDF's TCP at the recorded joints, else the rig profile's actual_pose_lead_ms: sidecar conversion.eef_timing." }, "eef_state": { "feature": "observation.eef_state", "meaning": "every arm's measured TCP pose (observation.eef_pose., sign-continuous) and gripper opening (observation.state's) in one vector, the measured counterpart of a tcp action: the state of an end-effector policy", "layout_per_arm": [ "x", "y", "z", "qx", "qy", "qz", "qw", "gripper" ], "arm_order": [ "left", "right" ], "gripper": "as observation.state's gripper; 0 for an arm without one" }, "kinematics": { "sidecar": "conversion.kinematics. of meta/paddy/episodes/.json", "tcp_vs_urdf": "tcp_vs_urdf_mm and tcp_vs_urdf_deg (median, max): the URDF's base_T_tcp at the recorded joints against the recorded TCP pose", "dh": "dh: the arm control box's own DH table (metadata.json controller_settings..dh_params: 28 values, four per joint, mm and rad) as identified against the recorded TCP pose: layout (the name of each of the four values per joint, in the table's order), convention (modified = Craig, or standard), tcp_vs_dh_mm and tcp_vs_dh_deg (median, max), samples, fit, and dh_vs_urdf_mm (the table's flange against the URDF's at the same joints); with fit true, forward kinematics of the table read that way is the arm's calibrated kinematics. {note} when the trial has no table" }, "cameras": { "file": "meta/cameras.json", "role": "where the camera is mounted, as the rig profile declares it: overhead (fixed, looking down at the workspace), scene (fixed, another viewpoint) or wrist (moves with an arm; which arm is mount.side); null = not declared", "primary": "the file's top-level primary names the rig's one primary camera (the view to look at first), and that camera's entry has primary: true; null = not declared", "base_T_cam": "4x4 pose of the camera's color optical frame (z forward, x right, y down) in each arm base frame, metres; extrinsics..source says where it came from", "sources": { "tf_static": "the nominal pose: the rig's URDF and the camera mounts in the rig profile (CAD or tape), as published on /tf_static", "tf_snapshot": "the nominal pose from the first /tf snapshot at episode start (valid: episode_start)", "urdf_fk": "the nominal pose of a wrist camera by forward kinematics of the recorded URDF and joint states at t_s, the episode's first slot (valid: episode_start)", "measured": "the rig profile's measured calibration: calibration.frame_T_cam chained to the base through via (identity, tf_static, urdf, urdf_fk at t_s for a wrist camera, or tf_snapshot); the nominal entry is kept as nominal {base_T_cam, source}" }, "measured_vs_nominal": "measured_vs_nominal_mm / _deg: distance of the origins and angle of the rotation between the measured and the nominal base_T_cam at the episode's first slot", "intrinsics": "K (row-major 3x3) and D (plumb_bob k1 k2 p1 p2 k3) describe the raw color image as recorded (not undistorted). intrinsics.source camera_info: the camera's own CameraInfo (a RealSense's factory values; a D435 color stream reports D all zeros). measured: the rig profile's measured intrinsics (method, calibrated_at, error_px, samples, notes, from), the CameraInfo values kept as factory_intrinsics {K, D, distortion_model}. The pose of a fixed camera measured after its intrinsics were is solved with them", "calibration": "per camera: the measured calibration (frame, xyz_mm, rpy_deg URDF fixed-axis, frame_T_cam, method, calibrated_at, error_mm, error_px, samples, notes, from: rig_snapshot | rig_option), null when the rig profile has none. TF on the rig stays nominal", "image_rotation": "image_rotation_deg per camera: 0, or 180 for a camera mounted upside down whose pictures the dataset turns. Turned, the color and depth frames are the recorded ones turned by 180 degrees and every value that describes them is of the turned image: K (cx' = W - 1 - cx, cy' = H - 1 - cy), D (p1 and p2 negated), base_T_cam, mount link_T_cam and tcp_T_cam, calibration.frame_T_cam, the depth K, D and color_T_depth, and paddy.camera_pose.: the camera's own optical frame (frame_id) turned about its z axis, X_T_turned = X_T_cam @ diag(-1, -1, 1, 1). calibration.xyz_mm and rpy_deg stay the rig profile's record of the camera's own optical frame. A policy trained on it needs the live frames turned the same way (paddy's policy server does it from this file)", "per_row_pose": "paddy.camera_pose. of a wrist camera; source measured_fk (forward kinematics to the calibration frame times its measured pose), urdf_fk or eef; time row (at the joint sample observation.state holds, so it matches observation.state through the URDF; the image is paddy.frame_age. older) or frame (at the shown frame's stamp)" }, "subtasks": { "feature": "paddy.subtask_index", "meaning": "the operator-marked subtask open at the row (metadata.json subtasks): its index, -1 where none is open. Spans are half-open [start, end) in header.stamp POSIX seconds; paused spans are out of the grid, so a subtask with a pause inside has no rows for the paused time", "names": "the episode sidecar's conversion.subtasks: per index the plain-language instruction, the outcome (success, failure or unknown), start_s, end_s (null: still open at the end of the recording) and the rows [start_row, end_row) it covers (null when none)", "not_a_policy_input": true }, "command_sent": { "feature": "paddy.command_sent.", "pose": "the pose the follower sent to the arm (its mm / Euler call converted back to m / quaternion xyzw in the base frame; for an xArm5 the projected pose actually sent), zero-order held with action_hold_s; the call's speed and acceleration are in the sidecar's conversion.motion_profile.", "before_first_command": "the measured TCP pose (observation.eef_pose.), paddy.action_issued false: the arm stays where it is until its first command. The TCP target (a tcp dataset's action, paddy.target_pose. in a joints dataset) holds the measured TCP pose before the first target too, paddy.action_valid false. Zeros only for an arm whose stream, or TCP pose, was never recorded" }, "velocity": { "feature": "observation.velocity", "status": "present", "unit": "rad/s", "source": "sdk", "meaning": "joint velocities the xArm SDK reports (get_joint_states), as the follower published them with each joint-state sample (100 Hz), zero-order held like observation.state", "order": "per arm joint1..jointN (URDF order, base to wrist), as observation.state; no gripper" }, "effort": { "feature": "observation.effort", "status": "present", "unit": "N m", "source": "sdk_estimate", "validated": false, "meaning": "the xArm SDK's joint torque estimate (get_joint_states) from motor current and the arm's own dynamic model; the arms have no joint torque sensors. Sign and magnitude not yet checked on the rig", "order": "per arm joint1..jointN (URDF order, base to wrist), as observation.state; no gripper", "validation": "static torques at four poses against the URDF masses' gravity torques, then the gripper pressed on a kitchen scale" }, "joint_state_status": "present: the column is written; absent: a bag's joint states carry none, so the dataset leaves it out (never zeros; each sidecar's conversion.joint_velocity / joint_effort names the episodes); off: turned off (--no-joint-velocity, --no-joint-effort)", "rot6d": "first_two_columns_column_major", "rot6d_features": { "layout": [ "r00", "r10", "r20", "r01", "r11", "r21" ], "meaning": "the first two columns of the rotation matrix R (base_T_tcp's rotation), column-major: [r00, r10, r20, r01, r11, r21] (Zhou et al. 2019; GR00T XYZ_ROT6D, OpenNeoData). Continuous by construction; recover R by Gram-Schmidt on the two columns and their cross product. Zeros where the quaternion is zero (a stream that was never recorded)", "observation": "observation.eef_rot6d.", "target": "paddy.target_rot6d.", "not_a_policy_output": "the target's 6D rotation is outside the action prefix" }, "force_torque": { "feature": "observation.force_torque.", "raw_feature": "paddy.force_torque_raw.", "valid_mask": "paddy.force_torque_valid", "arm_order": [ "right" ], "layout": [ "fx", "fy", "fz", "tx", "ty", "tz" ], "units": { "force": "N", "torque": "N m" }, "frames": { "right": "Right_link_eef" }, "frame_note": "the flange (URDF link_eef), which the six-axis sensor is mounted on. Checked on the rig (2026-11-07): a press along each flange axis reads on that axis with the sign of the force applied to the tool (a push toward flange +x reads +fx); a load fit put the sensor axes within 1.4 deg of the flange's", "reading": "observation.force_torque.: the control box's reading after its load and offset compensation (/paddy/_arm/ft_wrench, SDK ft_ext_force): near zero, up to a small offset, with nothing touching the tool when the payload set on the control box matches the tool. paddy.force_torque_raw.: the unprocessed reading (/paddy/_arm/ft_wrench_raw, SDK ft_raw_force), so the compensation can be redone", "compensation": "the payload the control box compensated for is in each episode sidecar's metadata.controller_settings..tcp_load_kg and tcp_load_cog_mm, read at the trial's start", "timing": "row k holds the last reading stamped at or before t_k (zero-order hold, fresh within hold_max_age_s); the stamp is when the follower's host read the control box's real-time report packet", "missing": "rows without a fresh reading hold zeros and paddy.force_torque_valid false; they are not dropped. An arm whose rig declares the sensor but recorded no readings has the columns, all zeros and invalid", "rate_hz_nominal": "about 100 (the real-time report); 5 when the follower fell back to the rich report", "measured_median_hz": { "right": 99.5 }, "measured_per_episode": "sidecar conversion.streams.force_torque..rate_hz; conversion.force_torque_valid_ratio", "hold_max_age_s": 0.2 }, "target_pose": { "feature": "paddy.target_pose.", "pose": "the last TCP target the teleop controller published (/paddy/_arm/target_pose): absolute in the arm base frame (action.frames, TCP action.tcp_frames), m and quaternion xyzw, sign-continuous per episode, held with action_hold_s; what action holds in a tcp dataset. Zeros for an arm driven by joint targets alone (a GELLO in joint mode)" } }, "timing": { "camera_sync_worst_median_ms": 10.634, "command_lag_ms": { "left": 113.925, "right": 117.21 }, "action_latency_ms": { "left": { "arm": 113.925, "gripper": 119.85 }, "right": { "arm": 117.21, "gripper": 111.15 } }, "action_latency_per_episode": "meta/paddy/episodes/{trial_key}.json conversion.action_latency_ms..arm / .gripper" }, "merge": { "version": 1, "merged_at": "2026-11-14T07:32:08.203862+00:00", "paddy_version": "0.2.0.dev0", "files": "hard-linked where possible", "stats": "aggregated from the per-episode statistics with LeRobot's aggregate_stats formula; quantiles are count-weighted means of the episodes' quantiles", "differing_keys": [ "control_rate_hz", "conventions", "timing" ], "median_over_sources": [ "control_rate_hz", "conventions", "timing" ], "sources": [ { "path": "policy/datasets/cablesort_2026-11-13_parts_showcase_v2/2026-11-14_00-40-33_trial_007", "trials": [ "trial_007" ], "episode_keys": [ "bf65b2a0-5809-412d-8aef-8b8fa5c07502" ], "sessions": [ "session_2026-11-14_00-40-33" ], "episodes": 1, "frames": 1147, "control_rate_hz": { "left": 68.01096139190219, "right": 67.9217515222179 }, "conventions": { "version": 1, "action": { "feature": "action", "space": "joints", "meaning": "per arm the joint command, zero-order held, then the gripper command: the quantity a policy that drives the arms through target_joints outputs. The TCP target the teleop controller published is paddy.target_pose.; the pose the follower sent is paddy.command_sent.", "joints": "absolute positions of the URDF joints joint1..jointN, base to wrist, in observation.state's order; not a delta, not a TCP pose", "joint_unit": "rad", "joints_source": { "left": "command_sent", "right": "command_sent" }, "joints_sources": { "command_sent": "the joint angles the follower sent (command_sent joints_rad: a set_servo_angle call's target or, in servo mode, a set_servo_angle_j setpoint; the xArm7 joint path, GELLO and policy arms), held with action_hold_s; timing.command_methods says which", "target_joints": "a leader arm's recorded joint targets (target_joints, GELLO), held with action_hold_s", "next_state": "the next row's measured joints (the last row its own): the usual stand-in for a position-controlled arm's joint command when none was recorded (a mode-7 arm, older bags)" }, "joints_source_per_episode": "conversion.action_source. of meta/paddy/episodes/.json", "before_first_command": "the row's measured joints (observation.state), paddy.action_valid false: the arm stays where it is until its first command, as the gripper command holds the measured opening", "layout_per_arm": { "left": [ "joint1", "joint2", "joint3", "joint4", "joint5", "joint6", "joint7", "gripper_cmd" ], "right": [ "joint1", "joint2", "joint3", "joint4", "joint5", "joint6", "joint7", "gripper_cmd" ] }, "frames": { "left": "Left_link_base", "right": "Right_link_base" }, "tcp_frames": { "left": "Left_link_tcp", "right": "Right_link_tcp" }, "arm_order": [ "left", "right" ], "gripper": { "unit": "fraction, 0 closed .. 1 open", "physical": { "left": { "model": "g2", "source": "metadata", "state_closed": 0.84, "state_open": 0.0, "cmd_closed": -10.0, "cmd_open": 850.0 }, "right": { "model": "g2", "source": "metadata", "state_closed": 0.84, "state_open": 0.0, "cmd_closed": -10.0, "cmd_open": 850.0 } } }, "already_scaled": true, "scaling_note": "action is in robot units after the controller's motion scaling and rate limit; paddy.scaling is the human-input gain in force (informational, not a policy input)" }, "timing": { "row_time": "t_k = t0 + k / fps (t0 = the latest first stamp over the arms' joint states and the cameras' color frames; paused spans removed); timestamp = k / fps; paddy.utc_s and the sidecar's conversion.t_start_s give the absolute time", "rule": "row k holds the last joint command stamped at or before t_k (zero-order hold, no shift): the command issued at t_k; with joints_source next_state the measured joints of row k+1, the position the arm reached next", "stamp_meaning": "command_sent: header.stamp of the command_sent message, when the follower's motion call (set_servo_angle, or set_servo_angle_j in servo mode) returned; target_joints: header.stamp of the leader's target; next_state: the next row's slot time", "execution": "the follower validates each target and commands the arm from it; how depends on the arm's control mode, named per command by command_sent.method (per arm command_methods, per episode the sidecar's conversion.streams.command_sent..methods) and described in execution_by_method. The issue time of each command is header.stamp of the command_sent stream", "command_methods": { "left": [ "set_servo_angle_j" ], "right": [ "set_servo_angle_j" ] }, "execution_by_method": { "set_servo_angle_j": "servo mode (xArm mode 1): the follower takes each accepted target as joints (its own solution of a TCP target, or a leader's joint target), steps a joint setpoint toward the newest one on a fixed tick (rates.control.servo_rate_hz), each joint inside the stream's speed and acceleration caps, and sends every setpoint with set_servo_angle_j; the control box executes the last setpoint it received and plans nothing. command_sent has one entry per setpoint: joints_rad is the setpoint, the pose and target_stamp are those of the target it steps toward" }, "action_hold_s": 0.2, "valid_mask": "paddy.action_valid", "issued_mask": "paddy.action_issued" }, "rates": { "video_fps": 30, "control": { "teleop_target_hz_nominal": "50 with the target buffer on, else the Quest sample rate (about 50 to 72)", "target_buffer_ms": { "left": 0.0, "right": 0.0 }, "servo_rate_hz": { "left": 200.0, "right": 200.0 }, "measured_median_hz": { "action_joints.left": 198.79160149770132, "action_joints.right": 228.7906177553525, "command_sent.left": 198.79160149770132, "command_sent.right": 228.7906177553525, "target.left": 68.01096139190219, "target.right": 67.9217515222179 }, "measured_per_episode": "sidecar conversion.streams.target..rate_hz and conversion.streams.command_sent..rate_hz" }, "observation_state_hz": 100, "eef_pose_hz": 100, "safety_state_hz": 10 }, "observation_state": { "joint_unit": "rad", "joint_order": "per arm the URDF joints joint1..jointN, base to wrist, then the gripper fraction; arms in arm_order", "gripper_unit": "fraction, 0 closed .. 1 open", "gripper_opening_m": "per arm, meta/paddy/info.json grippers..opening_m: the opening between the jaws at fraction 0 and at fraction 1, metres; the fraction is linear in it" }, "eef_pose": { "feature": "observation.eef_pose.", "pose": "measured TCP pose in the arm base frame, m and quaternion xyzw, zero-order held", "mode": "urdf", "source": { "left": "urdf_fk", "right": "urdf_fk" }, "sources": "urdf_fk: the recorded URDF's base_T_tcp at the recorded joint states (100 Hz), the kinematics the wrist cameras' paddy.camera_pose and the teleop targets use; reported: the pose the arm's control box reports (about 50 Hz, its own kinematics, which can differ from the URDF's by several mm: sidecar conversion.kinematics..tcp_vs_urdf_mm); mixed: episodes differ, see each sidecar's conversion.tcp_pose.", "stamps": "urdf_fk: the joint states' stamps. reported: the control box reports the pose ahead of the arm; each sample is restamped by that lead, measured per episode against the URDF's TCP at the recorded joints, else the rig profile's actual_pose_lead_ms: sidecar conversion.eef_timing." }, "eef_state": { "feature": "observation.eef_state", "meaning": "every arm's measured TCP pose (observation.eef_pose., sign-continuous) and gripper opening (observation.state's) in one vector, the measured counterpart of a tcp action: the state of an end-effector policy", "layout_per_arm": [ "x", "y", "z", "qx", "qy", "qz", "qw", "gripper" ], "arm_order": [ "left", "right" ], "gripper": "as observation.state's gripper; 0 for an arm without one" }, "kinematics": { "sidecar": "conversion.kinematics. of meta/paddy/episodes/.json", "tcp_vs_urdf": "tcp_vs_urdf_mm and tcp_vs_urdf_deg (median, max): the URDF's base_T_tcp at the recorded joints against the recorded TCP pose", "dh": "dh: the arm control box's own DH table (metadata.json controller_settings..dh_params: 28 values, four per joint, mm and rad) as identified against the recorded TCP pose: layout (the name of each of the four values per joint, in the table's order), convention (modified = Craig, or standard), tcp_vs_dh_mm and tcp_vs_dh_deg (median, max), samples, fit, and dh_vs_urdf_mm (the table's flange against the URDF's at the same joints); with fit true, forward kinematics of the table read that way is the arm's calibrated kinematics. {note} when the trial has no table" }, "cameras": { "file": "meta/cameras.json", "role": "where the camera is mounted, as the rig profile declares it: overhead (fixed, looking down at the workspace), scene (fixed, another viewpoint) or wrist (moves with an arm; which arm is mount.side); null = not declared", "primary": "the file's top-level primary names the rig's one primary camera (the view to look at first), and that camera's entry has primary: true; null = not declared", "base_T_cam": "4x4 pose of the camera's color optical frame (z forward, x right, y down) in each arm base frame, metres; extrinsics..source says where it came from", "sources": { "tf_static": "the nominal pose: the rig's URDF and the camera mounts in the rig profile (CAD or tape), as published on /tf_static", "tf_snapshot": "the nominal pose from the first /tf snapshot at episode start (valid: episode_start)", "urdf_fk": "the nominal pose of a wrist camera by forward kinematics of the recorded URDF and joint states at t_s, the episode's first slot (valid: episode_start)", "measured": "the rig profile's measured calibration: calibration.frame_T_cam chained to the base through via (identity, tf_static, urdf, urdf_fk at t_s for a wrist camera, or tf_snapshot); the nominal entry is kept as nominal {base_T_cam, source}" }, "measured_vs_nominal": "measured_vs_nominal_mm / _deg: distance of the origins and angle of the rotation between the measured and the nominal base_T_cam at the episode's first slot", "intrinsics": "K (row-major 3x3) and D (plumb_bob k1 k2 p1 p2 k3) describe the raw color image as recorded (not undistorted). intrinsics.source camera_info: the camera's own CameraInfo (a RealSense's factory values; a D435 color stream reports D all zeros). measured: the rig profile's measured intrinsics (method, calibrated_at, error_px, samples, notes, from), the CameraInfo values kept as factory_intrinsics {K, D, distortion_model}. The pose of a fixed camera measured after its intrinsics were is solved with them", "calibration": "per camera: the measured calibration (frame, xyz_mm, rpy_deg URDF fixed-axis, frame_T_cam, method, calibrated_at, error_mm, error_px, samples, notes, from: rig_snapshot | rig_option), null when the rig profile has none. TF on the rig stays nominal", "image_rotation": "image_rotation_deg per camera: 0, or 180 for a camera mounted upside down whose pictures the dataset turns. Turned, the color and depth frames are the recorded ones turned by 180 degrees and every value that describes them is of the turned image: K (cx' = W - 1 - cx, cy' = H - 1 - cy), D (p1 and p2 negated), base_T_cam, mount link_T_cam and tcp_T_cam, calibration.frame_T_cam, the depth K, D and color_T_depth, and paddy.camera_pose.: the camera's own optical frame (frame_id) turned about its z axis, X_T_turned = X_T_cam @ diag(-1, -1, 1, 1). calibration.xyz_mm and rpy_deg stay the rig profile's record of the camera's own optical frame. A policy trained on it needs the live frames turned the same way (paddy's policy server does it from this file)", "per_row_pose": "paddy.camera_pose. of a wrist camera; source measured_fk (forward kinematics to the calibration frame times its measured pose), urdf_fk or eef; time row (at the joint sample observation.state holds, so it matches observation.state through the URDF; the image is paddy.frame_age. older) or frame (at the shown frame's stamp)" }, "subtasks": { "feature": "paddy.subtask_index", "meaning": "the operator-marked subtask open at the row (metadata.json subtasks): its index, -1 where none is open. Spans are half-open [start, end) in header.stamp POSIX seconds; paused spans are out of the grid, so a subtask with a pause inside has no rows for the paused time", "names": "the episode sidecar's conversion.subtasks: per index the plain-language instruction, the outcome (success, failure or unknown), start_s, end_s (null: still open at the end of the recording) and the rows [start_row, end_row) it covers (null when none)", "not_a_policy_input": true }, "command_sent": { "feature": "paddy.command_sent.", "pose": "the pose the follower sent to the arm (its mm / Euler call converted back to m / quaternion xyzw in the base frame; for an xArm5 the projected pose actually sent), zero-order held with action_hold_s; the call's speed and acceleration are in the sidecar's conversion.motion_profile.", "before_first_command": "the measured TCP pose (observation.eef_pose.), paddy.action_issued false: the arm stays where it is until its first command. The TCP target (a tcp dataset's action, paddy.target_pose. in a joints dataset) holds the measured TCP pose before the first target too, paddy.action_valid false. Zeros only for an arm whose stream, or TCP pose, was never recorded" }, "velocity": { "feature": "observation.velocity", "status": "present", "unit": "rad/s", "source": "sdk", "meaning": "joint velocities the xArm SDK reports (get_joint_states), as the follower published them with each joint-state sample (100 Hz), zero-order held like observation.state", "order": "per arm joint1..jointN (URDF order, base to wrist), as observation.state; no gripper" }, "effort": { "feature": "observation.effort", "status": "present", "unit": "N m", "source": "sdk_estimate", "validated": false, "meaning": "the xArm SDK's joint torque estimate (get_joint_states) from motor current and the arm's own dynamic model; the arms have no joint torque sensors. Sign and magnitude not yet checked on the rig", "order": "per arm joint1..jointN (URDF order, base to wrist), as observation.state; no gripper", "validation": "static torques at four poses against the URDF masses' gravity torques, then the gripper pressed on a kitchen scale" }, "joint_state_status": "present: the column is written; absent: a bag's joint states carry none, so the dataset leaves it out (never zeros; each sidecar's conversion.joint_velocity / joint_effort names the episodes); off: turned off (--no-joint-velocity, --no-joint-effort)", "rot6d": "first_two_columns_column_major", "rot6d_features": { "layout": [ "r00", "r10", "r20", "r01", "r11", "r21" ], "meaning": "the first two columns of the rotation matrix R (base_T_tcp's rotation), column-major: [r00, r10, r20, r01, r11, r21] (Zhou et al. 2019; GR00T XYZ_ROT6D, OpenNeoData). Continuous by construction; recover R by Gram-Schmidt on the two columns and their cross product. Zeros where the quaternion is zero (a stream that was never recorded)", "observation": "observation.eef_rot6d.", "target": "paddy.target_rot6d.", "not_a_policy_output": "the target's 6D rotation is outside the action prefix" }, "force_torque": { "feature": "observation.force_torque.", "raw_feature": "paddy.force_torque_raw.", "valid_mask": "paddy.force_torque_valid", "arm_order": [ "right" ], "layout": [ "fx", "fy", "fz", "tx", "ty", "tz" ], "units": { "force": "N", "torque": "N m" }, "frames": { "right": "Right_link_eef" }, "frame_note": "the flange (URDF link_eef), which the six-axis sensor is mounted on. Checked on the rig (2026-11-07): a press along each flange axis reads on that axis with the sign of the force applied to the tool (a push toward flange +x reads +fx); a load fit put the sensor axes within 1.4 deg of the flange's", "reading": "observation.force_torque.: the control box's reading after its load and offset compensation (/paddy/_arm/ft_wrench, SDK ft_ext_force): near zero, up to a small offset, with nothing touching the tool when the payload set on the control box matches the tool. paddy.force_torque_raw.: the unprocessed reading (/paddy/_arm/ft_wrench_raw, SDK ft_raw_force), so the compensation can be redone", "compensation": "the payload the control box compensated for is in each episode sidecar's metadata.controller_settings..tcp_load_kg and tcp_load_cog_mm, read at the trial's start", "timing": "row k holds the last reading stamped at or before t_k (zero-order hold, fresh within hold_max_age_s); the stamp is when the follower's host read the control box's real-time report packet", "missing": "rows without a fresh reading hold zeros and paddy.force_torque_valid false; they are not dropped. An arm whose rig declares the sensor but recorded no readings has the columns, all zeros and invalid", "rate_hz_nominal": "about 100 (the real-time report); 5 when the follower fell back to the rich report", "measured_median_hz": { "right": 99.5 }, "measured_per_episode": "sidecar conversion.streams.force_torque..rate_hz; conversion.force_torque_valid_ratio", "hold_max_age_s": 0.2 }, "target_pose": { "feature": "paddy.target_pose.", "pose": "the last TCP target the teleop controller published (/paddy/_arm/target_pose): absolute in the arm base frame (action.frames, TCP action.tcp_frames), m and quaternion xyzw, sign-continuous per episode, held with action_hold_s; what action holds in a tcp dataset. Zeros for an arm driven by joint targets alone (a GELLO in joint mode)" } }, "timing": { "camera_sync_worst_median_ms": 8.768, "command_lag_ms": { "left": 113.67, "right": 117.43 }, "action_latency_ms": { "left": { "arm": 113.67, "gripper": 129.4 }, "right": { "arm": 117.43, "gripper": 106.6 } }, "action_latency_per_episode": "meta/paddy/episodes/{trial_key}.json conversion.action_latency_ms..arm / .gripper" } }, { "path": "policy/datasets/cablesort_2026-11-13_parts_showcase_v2/2026-11-14_00-40-33_trial_010", "trials": [ "trial_010" ], "episode_keys": [ "f862d332-f27a-4944-96e1-5e1d42a21b58" ], "sessions": [ "session_2026-11-14_00-40-33" ], "episodes": 1, "frames": 983, "control_rate_hz": { "left": 67.76811219543721, "right": 69.12168754119973 }, "conventions": { "version": 1, "action": { "feature": "action", "space": "joints", "meaning": "per arm the joint command, zero-order held, then the gripper command: the quantity a policy that drives the arms through target_joints outputs. The TCP target the teleop controller published is paddy.target_pose.; the pose the follower sent is paddy.command_sent.", "joints": "absolute positions of the URDF joints joint1..jointN, base to wrist, in observation.state's order; not a delta, not a TCP pose", "joint_unit": "rad", "joints_source": { "left": "command_sent", "right": "command_sent" }, "joints_sources": { "command_sent": "the joint angles the follower sent (command_sent joints_rad: a set_servo_angle call's target or, in servo mode, a set_servo_angle_j setpoint; the xArm7 joint path, GELLO and policy arms), held with action_hold_s; timing.command_methods says which", "target_joints": "a leader arm's recorded joint targets (target_joints, GELLO), held with action_hold_s", "next_state": "the next row's measured joints (the last row its own): the usual stand-in for a position-controlled arm's joint command when none was recorded (a mode-7 arm, older bags)" }, "joints_source_per_episode": "conversion.action_source. of meta/paddy/episodes/.json", "before_first_command": "the row's measured joints (observation.state), paddy.action_valid false: the arm stays where it is until its first command, as the gripper command holds the measured opening", "layout_per_arm": { "left": [ "joint1", "joint2", "joint3", "joint4", "joint5", "joint6", "joint7", "gripper_cmd" ], "right": [ "joint1", "joint2", "joint3", "joint4", "joint5", "joint6", "joint7", "gripper_cmd" ] }, "frames": { "left": "Left_link_base", "right": "Right_link_base" }, "tcp_frames": { "left": "Left_link_tcp", "right": "Right_link_tcp" }, "arm_order": [ "left", "right" ], "gripper": { "unit": "fraction, 0 closed .. 1 open", "physical": { "left": { "model": "g2", "source": "metadata", "state_closed": 0.84, "state_open": 0.0, "cmd_closed": -10.0, "cmd_open": 850.0 }, "right": { "model": "g2", "source": "metadata", "state_closed": 0.84, "state_open": 0.0, "cmd_closed": -10.0, "cmd_open": 850.0 } } }, "already_scaled": true, "scaling_note": "action is in robot units after the controller's motion scaling and rate limit; paddy.scaling is the human-input gain in force (informational, not a policy input)" }, "timing": { "row_time": "t_k = t0 + k / fps (t0 = the latest first stamp over the arms' joint states and the cameras' color frames; paused spans removed); timestamp = k / fps; paddy.utc_s and the sidecar's conversion.t_start_s give the absolute time", "rule": "row k holds the last joint command stamped at or before t_k (zero-order hold, no shift): the command issued at t_k; with joints_source next_state the measured joints of row k+1, the position the arm reached next", "stamp_meaning": "command_sent: header.stamp of the command_sent message, when the follower's motion call (set_servo_angle, or set_servo_angle_j in servo mode) returned; target_joints: header.stamp of the leader's target; next_state: the next row's slot time", "execution": "the follower validates each target and commands the arm from it; how depends on the arm's control mode, named per command by command_sent.method (per arm command_methods, per episode the sidecar's conversion.streams.command_sent..methods) and described in execution_by_method. The issue time of each command is header.stamp of the command_sent stream", "command_methods": { "left": [ "set_servo_angle_j" ], "right": [ "set_servo_angle_j" ] }, "execution_by_method": { "set_servo_angle_j": "servo mode (xArm mode 1): the follower takes each accepted target as joints (its own solution of a TCP target, or a leader's joint target), steps a joint setpoint toward the newest one on a fixed tick (rates.control.servo_rate_hz), each joint inside the stream's speed and acceleration caps, and sends every setpoint with set_servo_angle_j; the control box executes the last setpoint it received and plans nothing. command_sent has one entry per setpoint: joints_rad is the setpoint, the pose and target_stamp are those of the target it steps toward" }, "action_hold_s": 0.2, "valid_mask": "paddy.action_valid", "issued_mask": "paddy.action_issued" }, "rates": { "video_fps": 30, "control": { "teleop_target_hz_nominal": "50 with the target buffer on, else the Quest sample rate (about 50 to 72)", "target_buffer_ms": { "left": 0.0, "right": 0.0 }, "servo_rate_hz": { "left": 200.0, "right": 200.0 }, "measured_median_hz": { "action_joints.left": 198.98021727785948, "action_joints.right": 242.50138760407032, "command_sent.left": 198.98021727785948, "command_sent.right": 242.50138760407032, "target.left": 67.76811219543721, "target.right": 69.12168754119973 }, "measured_per_episode": "sidecar conversion.streams.target..rate_hz and conversion.streams.command_sent..rate_hz" }, "observation_state_hz": 100, "eef_pose_hz": 100, "safety_state_hz": 10 }, "observation_state": { "joint_unit": "rad", "joint_order": "per arm the URDF joints joint1..jointN, base to wrist, then the gripper fraction; arms in arm_order", "gripper_unit": "fraction, 0 closed .. 1 open", "gripper_opening_m": "per arm, meta/paddy/info.json grippers..opening_m: the opening between the jaws at fraction 0 and at fraction 1, metres; the fraction is linear in it" }, "eef_pose": { "feature": "observation.eef_pose.", "pose": "measured TCP pose in the arm base frame, m and quaternion xyzw, zero-order held", "mode": "urdf", "source": { "left": "urdf_fk", "right": "urdf_fk" }, "sources": "urdf_fk: the recorded URDF's base_T_tcp at the recorded joint states (100 Hz), the kinematics the wrist cameras' paddy.camera_pose and the teleop targets use; reported: the pose the arm's control box reports (about 50 Hz, its own kinematics, which can differ from the URDF's by several mm: sidecar conversion.kinematics..tcp_vs_urdf_mm); mixed: episodes differ, see each sidecar's conversion.tcp_pose.", "stamps": "urdf_fk: the joint states' stamps. reported: the control box reports the pose ahead of the arm; each sample is restamped by that lead, measured per episode against the URDF's TCP at the recorded joints, else the rig profile's actual_pose_lead_ms: sidecar conversion.eef_timing." }, "eef_state": { "feature": "observation.eef_state", "meaning": "every arm's measured TCP pose (observation.eef_pose., sign-continuous) and gripper opening (observation.state's) in one vector, the measured counterpart of a tcp action: the state of an end-effector policy", "layout_per_arm": [ "x", "y", "z", "qx", "qy", "qz", "qw", "gripper" ], "arm_order": [ "left", "right" ], "gripper": "as observation.state's gripper; 0 for an arm without one" }, "kinematics": { "sidecar": "conversion.kinematics. of meta/paddy/episodes/.json", "tcp_vs_urdf": "tcp_vs_urdf_mm and tcp_vs_urdf_deg (median, max): the URDF's base_T_tcp at the recorded joints against the recorded TCP pose", "dh": "dh: the arm control box's own DH table (metadata.json controller_settings..dh_params: 28 values, four per joint, mm and rad) as identified against the recorded TCP pose: layout (the name of each of the four values per joint, in the table's order), convention (modified = Craig, or standard), tcp_vs_dh_mm and tcp_vs_dh_deg (median, max), samples, fit, and dh_vs_urdf_mm (the table's flange against the URDF's at the same joints); with fit true, forward kinematics of the table read that way is the arm's calibrated kinematics. {note} when the trial has no table" }, "cameras": { "file": "meta/cameras.json", "role": "where the camera is mounted, as the rig profile declares it: overhead (fixed, looking down at the workspace), scene (fixed, another viewpoint) or wrist (moves with an arm; which arm is mount.side); null = not declared", "primary": "the file's top-level primary names the rig's one primary camera (the view to look at first), and that camera's entry has primary: true; null = not declared", "base_T_cam": "4x4 pose of the camera's color optical frame (z forward, x right, y down) in each arm base frame, metres; extrinsics..source says where it came from", "sources": { "tf_static": "the nominal pose: the rig's URDF and the camera mounts in the rig profile (CAD or tape), as published on /tf_static", "tf_snapshot": "the nominal pose from the first /tf snapshot at episode start (valid: episode_start)", "urdf_fk": "the nominal pose of a wrist camera by forward kinematics of the recorded URDF and joint states at t_s, the episode's first slot (valid: episode_start)", "measured": "the rig profile's measured calibration: calibration.frame_T_cam chained to the base through via (identity, tf_static, urdf, urdf_fk at t_s for a wrist camera, or tf_snapshot); the nominal entry is kept as nominal {base_T_cam, source}" }, "measured_vs_nominal": "measured_vs_nominal_mm / _deg: distance of the origins and angle of the rotation between the measured and the nominal base_T_cam at the episode's first slot", "intrinsics": "K (row-major 3x3) and D (plumb_bob k1 k2 p1 p2 k3) describe the raw color image as recorded (not undistorted). intrinsics.source camera_info: the camera's own CameraInfo (a RealSense's factory values; a D435 color stream reports D all zeros). measured: the rig profile's measured intrinsics (method, calibrated_at, error_px, samples, notes, from), the CameraInfo values kept as factory_intrinsics {K, D, distortion_model}. The pose of a fixed camera measured after its intrinsics were is solved with them", "calibration": "per camera: the measured calibration (frame, xyz_mm, rpy_deg URDF fixed-axis, frame_T_cam, method, calibrated_at, error_mm, error_px, samples, notes, from: rig_snapshot | rig_option), null when the rig profile has none. TF on the rig stays nominal", "image_rotation": "image_rotation_deg per camera: 0, or 180 for a camera mounted upside down whose pictures the dataset turns. Turned, the color and depth frames are the recorded ones turned by 180 degrees and every value that describes them is of the turned image: K (cx' = W - 1 - cx, cy' = H - 1 - cy), D (p1 and p2 negated), base_T_cam, mount link_T_cam and tcp_T_cam, calibration.frame_T_cam, the depth K, D and color_T_depth, and paddy.camera_pose.: the camera's own optical frame (frame_id) turned about its z axis, X_T_turned = X_T_cam @ diag(-1, -1, 1, 1). calibration.xyz_mm and rpy_deg stay the rig profile's record of the camera's own optical frame. A policy trained on it needs the live frames turned the same way (paddy's policy server does it from this file)", "per_row_pose": "paddy.camera_pose. of a wrist camera; source measured_fk (forward kinematics to the calibration frame times its measured pose), urdf_fk or eef; time row (at the joint sample observation.state holds, so it matches observation.state through the URDF; the image is paddy.frame_age. older) or frame (at the shown frame's stamp)" }, "subtasks": { "feature": "paddy.subtask_index", "meaning": "the operator-marked subtask open at the row (metadata.json subtasks): its index, -1 where none is open. Spans are half-open [start, end) in header.stamp POSIX seconds; paused spans are out of the grid, so a subtask with a pause inside has no rows for the paused time", "names": "the episode sidecar's conversion.subtasks: per index the plain-language instruction, the outcome (success, failure or unknown), start_s, end_s (null: still open at the end of the recording) and the rows [start_row, end_row) it covers (null when none)", "not_a_policy_input": true }, "command_sent": { "feature": "paddy.command_sent.", "pose": "the pose the follower sent to the arm (its mm / Euler call converted back to m / quaternion xyzw in the base frame; for an xArm5 the projected pose actually sent), zero-order held with action_hold_s; the call's speed and acceleration are in the sidecar's conversion.motion_profile.", "before_first_command": "the measured TCP pose (observation.eef_pose.), paddy.action_issued false: the arm stays where it is until its first command. The TCP target (a tcp dataset's action, paddy.target_pose. in a joints dataset) holds the measured TCP pose before the first target too, paddy.action_valid false. Zeros only for an arm whose stream, or TCP pose, was never recorded" }, "velocity": { "feature": "observation.velocity", "status": "present", "unit": "rad/s", "source": "sdk", "meaning": "joint velocities the xArm SDK reports (get_joint_states), as the follower published them with each joint-state sample (100 Hz), zero-order held like observation.state", "order": "per arm joint1..jointN (URDF order, base to wrist), as observation.state; no gripper" }, "effort": { "feature": "observation.effort", "status": "present", "unit": "N m", "source": "sdk_estimate", "validated": false, "meaning": "the xArm SDK's joint torque estimate (get_joint_states) from motor current and the arm's own dynamic model; the arms have no joint torque sensors. Sign and magnitude not yet checked on the rig", "order": "per arm joint1..jointN (URDF order, base to wrist), as observation.state; no gripper", "validation": "static torques at four poses against the URDF masses' gravity torques, then the gripper pressed on a kitchen scale" }, "joint_state_status": "present: the column is written; absent: a bag's joint states carry none, so the dataset leaves it out (never zeros; each sidecar's conversion.joint_velocity / joint_effort names the episodes); off: turned off (--no-joint-velocity, --no-joint-effort)", "rot6d": "first_two_columns_column_major", "rot6d_features": { "layout": [ "r00", "r10", "r20", "r01", "r11", "r21" ], "meaning": "the first two columns of the rotation matrix R (base_T_tcp's rotation), column-major: [r00, r10, r20, r01, r11, r21] (Zhou et al. 2019; GR00T XYZ_ROT6D, OpenNeoData). Continuous by construction; recover R by Gram-Schmidt on the two columns and their cross product. Zeros where the quaternion is zero (a stream that was never recorded)", "observation": "observation.eef_rot6d.", "target": "paddy.target_rot6d.", "not_a_policy_output": "the target's 6D rotation is outside the action prefix" }, "force_torque": { "feature": "observation.force_torque.", "raw_feature": "paddy.force_torque_raw.", "valid_mask": "paddy.force_torque_valid", "arm_order": [ "right" ], "layout": [ "fx", "fy", "fz", "tx", "ty", "tz" ], "units": { "force": "N", "torque": "N m" }, "frames": { "right": "Right_link_eef" }, "frame_note": "the flange (URDF link_eef), which the six-axis sensor is mounted on. Checked on the rig (2026-11-07): a press along each flange axis reads on that axis with the sign of the force applied to the tool (a push toward flange +x reads +fx); a load fit put the sensor axes within 1.4 deg of the flange's", "reading": "observation.force_torque.: the control box's reading after its load and offset compensation (/paddy/_arm/ft_wrench, SDK ft_ext_force): near zero, up to a small offset, with nothing touching the tool when the payload set on the control box matches the tool. paddy.force_torque_raw.: the unprocessed reading (/paddy/_arm/ft_wrench_raw, SDK ft_raw_force), so the compensation can be redone", "compensation": "the payload the control box compensated for is in each episode sidecar's metadata.controller_settings..tcp_load_kg and tcp_load_cog_mm, read at the trial's start", "timing": "row k holds the last reading stamped at or before t_k (zero-order hold, fresh within hold_max_age_s); the stamp is when the follower's host read the control box's real-time report packet", "missing": "rows without a fresh reading hold zeros and paddy.force_torque_valid false; they are not dropped. An arm whose rig declares the sensor but recorded no readings has the columns, all zeros and invalid", "rate_hz_nominal": "about 100 (the real-time report); 5 when the follower fell back to the rich report", "measured_median_hz": { "right": 99.5 }, "measured_per_episode": "sidecar conversion.streams.force_torque..rate_hz; conversion.force_torque_valid_ratio", "hold_max_age_s": 0.2 }, "target_pose": { "feature": "paddy.target_pose.", "pose": "the last TCP target the teleop controller published (/paddy/_arm/target_pose): absolute in the arm base frame (action.frames, TCP action.tcp_frames), m and quaternion xyzw, sign-continuous per episode, held with action_hold_s; what action holds in a tcp dataset. Zeros for an arm driven by joint targets alone (a GELLO in joint mode)" } }, "timing": { "camera_sync_worst_median_ms": 7.169, "command_lag_ms": { "left": 114.18, "right": 116.99 }, "action_latency_ms": { "left": { "arm": 114.18, "gripper": 109.0 }, "right": { "arm": 116.99, "gripper": 115.4 } }, "action_latency_per_episode": "meta/paddy/episodes/{trial_key}.json conversion.action_latency_ms..arm / .gripper" } }, { "path": "policy/datasets/cablesort_2026-11-13_parts_showcase_v2/2026-11-14_02-10-30_trial_001", "trials": [ "trial_001" ], "episode_keys": [ "adfc51df-e8f7-44ae-946e-982b646a87b6" ], "sessions": [ "session_2026-11-14_02-10-30" ], "episodes": 1, "frames": 1123, "control_rate_hz": { "left": 67.2708522121268, "right": 68.3629132812309 }, "conventions": { "version": 1, "action": { "feature": "action", "space": "joints", "meaning": "per arm the joint command, zero-order held, then the gripper command: the quantity a policy that drives the arms through target_joints outputs. The TCP target the teleop controller published is paddy.target_pose.; the pose the follower sent is paddy.command_sent.", "joints": "absolute positions of the URDF joints joint1..jointN, base to wrist, in observation.state's order; not a delta, not a TCP pose", "joint_unit": "rad", "joints_source": { "left": "command_sent", "right": "command_sent" }, "joints_sources": { "command_sent": "the joint angles the follower sent (command_sent joints_rad: a set_servo_angle call's target or, in servo mode, a set_servo_angle_j setpoint; the xArm7 joint path, GELLO and policy arms), held with action_hold_s; timing.command_methods says which", "target_joints": "a leader arm's recorded joint targets (target_joints, GELLO), held with action_hold_s", "next_state": "the next row's measured joints (the last row its own): the usual stand-in for a position-controlled arm's joint command when none was recorded (a mode-7 arm, older bags)" }, "joints_source_per_episode": "conversion.action_source. of meta/paddy/episodes/.json", "before_first_command": "the row's measured joints (observation.state), paddy.action_valid false: the arm stays where it is until its first command, as the gripper command holds the measured opening", "layout_per_arm": { "left": [ "joint1", "joint2", "joint3", "joint4", "joint5", "joint6", "joint7", "gripper_cmd" ], "right": [ "joint1", "joint2", "joint3", "joint4", "joint5", "joint6", "joint7", "gripper_cmd" ] }, "frames": { "left": "Left_link_base", "right": "Right_link_base" }, "tcp_frames": { "left": "Left_link_tcp", "right": "Right_link_tcp" }, "arm_order": [ "left", "right" ], "gripper": { "unit": "fraction, 0 closed .. 1 open", "physical": { "left": { "model": "g2", "source": "metadata", "state_closed": 0.84, "state_open": 0.0, "cmd_closed": -10.0, "cmd_open": 850.0 }, "right": { "model": "g2", "source": "metadata", "state_closed": 0.84, "state_open": 0.0, "cmd_closed": -10.0, "cmd_open": 850.0 } } }, "already_scaled": true, "scaling_note": "action is in robot units after the controller's motion scaling and rate limit; paddy.scaling is the human-input gain in force (informational, not a policy input)" }, "timing": { "row_time": "t_k = t0 + k / fps (t0 = the latest first stamp over the arms' joint states and the cameras' color frames; paused spans removed); timestamp = k / fps; paddy.utc_s and the sidecar's conversion.t_start_s give the absolute time", "rule": "row k holds the last joint command stamped at or before t_k (zero-order hold, no shift): the command issued at t_k; with joints_source next_state the measured joints of row k+1, the position the arm reached next", "stamp_meaning": "command_sent: header.stamp of the command_sent message, when the follower's motion call (set_servo_angle, or set_servo_angle_j in servo mode) returned; target_joints: header.stamp of the leader's target; next_state: the next row's slot time", "execution": "the follower validates each target and commands the arm from it; how depends on the arm's control mode, named per command by command_sent.method (per arm command_methods, per episode the sidecar's conversion.streams.command_sent..methods) and described in execution_by_method. The issue time of each command is header.stamp of the command_sent stream", "command_methods": { "left": [ "set_servo_angle_j" ], "right": [ "set_servo_angle_j" ] }, "execution_by_method": { "set_servo_angle_j": "servo mode (xArm mode 1): the follower takes each accepted target as joints (its own solution of a TCP target, or a leader's joint target), steps a joint setpoint toward the newest one on a fixed tick (rates.control.servo_rate_hz), each joint inside the stream's speed and acceleration caps, and sends every setpoint with set_servo_angle_j; the control box executes the last setpoint it received and plans nothing. command_sent has one entry per setpoint: joints_rad is the setpoint, the pose and target_stamp are those of the target it steps toward" }, "action_hold_s": 0.2, "valid_mask": "paddy.action_valid", "issued_mask": "paddy.action_issued" }, "rates": { "video_fps": 30, "control": { "teleop_target_hz_nominal": "50 with the target buffer on, else the Quest sample rate (about 50 to 72)", "target_buffer_ms": { "left": 0.0, "right": 0.0 }, "servo_rate_hz": { "left": 200.0, "right": 200.0 }, "measured_median_hz": { "action_joints.left": 199.33957511525117, "action_joints.right": 217.5017631196847, "command_sent.left": 199.33957511525117, "command_sent.right": 217.5017631196847, "target.left": 67.2708522121268, "target.right": 68.3629132812309 }, "measured_per_episode": "sidecar conversion.streams.target..rate_hz and conversion.streams.command_sent..rate_hz" }, "observation_state_hz": 100, "eef_pose_hz": 100, "safety_state_hz": 10 }, "observation_state": { "joint_unit": "rad", "joint_order": "per arm the URDF joints joint1..jointN, base to wrist, then the gripper fraction; arms in arm_order", "gripper_unit": "fraction, 0 closed .. 1 open", "gripper_opening_m": "per arm, meta/paddy/info.json grippers..opening_m: the opening between the jaws at fraction 0 and at fraction 1, metres; the fraction is linear in it" }, "eef_pose": { "feature": "observation.eef_pose.", "pose": "measured TCP pose in the arm base frame, m and quaternion xyzw, zero-order held", "mode": "urdf", "source": { "left": "urdf_fk", "right": "urdf_fk" }, "sources": "urdf_fk: the recorded URDF's base_T_tcp at the recorded joint states (100 Hz), the kinematics the wrist cameras' paddy.camera_pose and the teleop targets use; reported: the pose the arm's control box reports (about 50 Hz, its own kinematics, which can differ from the URDF's by several mm: sidecar conversion.kinematics..tcp_vs_urdf_mm); mixed: episodes differ, see each sidecar's conversion.tcp_pose.", "stamps": "urdf_fk: the joint states' stamps. reported: the control box reports the pose ahead of the arm; each sample is restamped by that lead, measured per episode against the URDF's TCP at the recorded joints, else the rig profile's actual_pose_lead_ms: sidecar conversion.eef_timing." }, "eef_state": { "feature": "observation.eef_state", "meaning": "every arm's measured TCP pose (observation.eef_pose., sign-continuous) and gripper opening (observation.state's) in one vector, the measured counterpart of a tcp action: the state of an end-effector policy", "layout_per_arm": [ "x", "y", "z", "qx", "qy", "qz", "qw", "gripper" ], "arm_order": [ "left", "right" ], "gripper": "as observation.state's gripper; 0 for an arm without one" }, "kinematics": { "sidecar": "conversion.kinematics. of meta/paddy/episodes/.json", "tcp_vs_urdf": "tcp_vs_urdf_mm and tcp_vs_urdf_deg (median, max): the URDF's base_T_tcp at the recorded joints against the recorded TCP pose", "dh": "dh: the arm control box's own DH table (metadata.json controller_settings..dh_params: 28 values, four per joint, mm and rad) as identified against the recorded TCP pose: layout (the name of each of the four values per joint, in the table's order), convention (modified = Craig, or standard), tcp_vs_dh_mm and tcp_vs_dh_deg (median, max), samples, fit, and dh_vs_urdf_mm (the table's flange against the URDF's at the same joints); with fit true, forward kinematics of the table read that way is the arm's calibrated kinematics. {note} when the trial has no table" }, "cameras": { "file": "meta/cameras.json", "role": "where the camera is mounted, as the rig profile declares it: overhead (fixed, looking down at the workspace), scene (fixed, another viewpoint) or wrist (moves with an arm; which arm is mount.side); null = not declared", "primary": "the file's top-level primary names the rig's one primary camera (the view to look at first), and that camera's entry has primary: true; null = not declared", "base_T_cam": "4x4 pose of the camera's color optical frame (z forward, x right, y down) in each arm base frame, metres; extrinsics..source says where it came from", "sources": { "tf_static": "the nominal pose: the rig's URDF and the camera mounts in the rig profile (CAD or tape), as published on /tf_static", "tf_snapshot": "the nominal pose from the first /tf snapshot at episode start (valid: episode_start)", "urdf_fk": "the nominal pose of a wrist camera by forward kinematics of the recorded URDF and joint states at t_s, the episode's first slot (valid: episode_start)", "measured": "the rig profile's measured calibration: calibration.frame_T_cam chained to the base through via (identity, tf_static, urdf, urdf_fk at t_s for a wrist camera, or tf_snapshot); the nominal entry is kept as nominal {base_T_cam, source}" }, "measured_vs_nominal": "measured_vs_nominal_mm / _deg: distance of the origins and angle of the rotation between the measured and the nominal base_T_cam at the episode's first slot", "intrinsics": "K (row-major 3x3) and D (plumb_bob k1 k2 p1 p2 k3) describe the raw color image as recorded (not undistorted). intrinsics.source camera_info: the camera's own CameraInfo (a RealSense's factory values; a D435 color stream reports D all zeros). measured: the rig profile's measured intrinsics (method, calibrated_at, error_px, samples, notes, from), the CameraInfo values kept as factory_intrinsics {K, D, distortion_model}. The pose of a fixed camera measured after its intrinsics were is solved with them", "calibration": "per camera: the measured calibration (frame, xyz_mm, rpy_deg URDF fixed-axis, frame_T_cam, method, calibrated_at, error_mm, error_px, samples, notes, from: rig_snapshot | rig_option), null when the rig profile has none. TF on the rig stays nominal", "image_rotation": "image_rotation_deg per camera: 0, or 180 for a camera mounted upside down whose pictures the dataset turns. Turned, the color and depth frames are the recorded ones turned by 180 degrees and every value that describes them is of the turned image: K (cx' = W - 1 - cx, cy' = H - 1 - cy), D (p1 and p2 negated), base_T_cam, mount link_T_cam and tcp_T_cam, calibration.frame_T_cam, the depth K, D and color_T_depth, and paddy.camera_pose.: the camera's own optical frame (frame_id) turned about its z axis, X_T_turned = X_T_cam @ diag(-1, -1, 1, 1). calibration.xyz_mm and rpy_deg stay the rig profile's record of the camera's own optical frame. A policy trained on it needs the live frames turned the same way (paddy's policy server does it from this file)", "per_row_pose": "paddy.camera_pose. of a wrist camera; source measured_fk (forward kinematics to the calibration frame times its measured pose), urdf_fk or eef; time row (at the joint sample observation.state holds, so it matches observation.state through the URDF; the image is paddy.frame_age. older) or frame (at the shown frame's stamp)" }, "subtasks": { "feature": "paddy.subtask_index", "meaning": "the operator-marked subtask open at the row (metadata.json subtasks): its index, -1 where none is open. Spans are half-open [start, end) in header.stamp POSIX seconds; paused spans are out of the grid, so a subtask with a pause inside has no rows for the paused time", "names": "the episode sidecar's conversion.subtasks: per index the plain-language instruction, the outcome (success, failure or unknown), start_s, end_s (null: still open at the end of the recording) and the rows [start_row, end_row) it covers (null when none)", "not_a_policy_input": true }, "command_sent": { "feature": "paddy.command_sent.", "pose": "the pose the follower sent to the arm (its mm / Euler call converted back to m / quaternion xyzw in the base frame; for an xArm5 the projected pose actually sent), zero-order held with action_hold_s; the call's speed and acceleration are in the sidecar's conversion.motion_profile.", "before_first_command": "the measured TCP pose (observation.eef_pose.), paddy.action_issued false: the arm stays where it is until its first command. The TCP target (a tcp dataset's action, paddy.target_pose. in a joints dataset) holds the measured TCP pose before the first target too, paddy.action_valid false. Zeros only for an arm whose stream, or TCP pose, was never recorded" }, "velocity": { "feature": "observation.velocity", "status": "present", "unit": "rad/s", "source": "sdk", "meaning": "joint velocities the xArm SDK reports (get_joint_states), as the follower published them with each joint-state sample (100 Hz), zero-order held like observation.state", "order": "per arm joint1..jointN (URDF order, base to wrist), as observation.state; no gripper" }, "effort": { "feature": "observation.effort", "status": "present", "unit": "N m", "source": "sdk_estimate", "validated": false, "meaning": "the xArm SDK's joint torque estimate (get_joint_states) from motor current and the arm's own dynamic model; the arms have no joint torque sensors. Sign and magnitude not yet checked on the rig", "order": "per arm joint1..jointN (URDF order, base to wrist), as observation.state; no gripper", "validation": "static torques at four poses against the URDF masses' gravity torques, then the gripper pressed on a kitchen scale" }, "joint_state_status": "present: the column is written; absent: a bag's joint states carry none, so the dataset leaves it out (never zeros; each sidecar's conversion.joint_velocity / joint_effort names the episodes); off: turned off (--no-joint-velocity, --no-joint-effort)", "rot6d": "first_two_columns_column_major", "rot6d_features": { "layout": [ "r00", "r10", "r20", "r01", "r11", "r21" ], "meaning": "the first two columns of the rotation matrix R (base_T_tcp's rotation), column-major: [r00, r10, r20, r01, r11, r21] (Zhou et al. 2019; GR00T XYZ_ROT6D, OpenNeoData). Continuous by construction; recover R by Gram-Schmidt on the two columns and their cross product. Zeros where the quaternion is zero (a stream that was never recorded)", "observation": "observation.eef_rot6d.", "target": "paddy.target_rot6d.", "not_a_policy_output": "the target's 6D rotation is outside the action prefix" }, "force_torque": { "feature": "observation.force_torque.", "raw_feature": "paddy.force_torque_raw.", "valid_mask": "paddy.force_torque_valid", "arm_order": [ "right" ], "layout": [ "fx", "fy", "fz", "tx", "ty", "tz" ], "units": { "force": "N", "torque": "N m" }, "frames": { "right": "Right_link_eef" }, "frame_note": "the flange (URDF link_eef), which the six-axis sensor is mounted on. Checked on the rig (2026-11-07): a press along each flange axis reads on that axis with the sign of the force applied to the tool (a push toward flange +x reads +fx); a load fit put the sensor axes within 1.4 deg of the flange's", "reading": "observation.force_torque.: the control box's reading after its load and offset compensation (/paddy/_arm/ft_wrench, SDK ft_ext_force): near zero, up to a small offset, with nothing touching the tool when the payload set on the control box matches the tool. paddy.force_torque_raw.: the unprocessed reading (/paddy/_arm/ft_wrench_raw, SDK ft_raw_force), so the compensation can be redone", "compensation": "the payload the control box compensated for is in each episode sidecar's metadata.controller_settings..tcp_load_kg and tcp_load_cog_mm, read at the trial's start", "timing": "row k holds the last reading stamped at or before t_k (zero-order hold, fresh within hold_max_age_s); the stamp is when the follower's host read the control box's real-time report packet", "missing": "rows without a fresh reading hold zeros and paddy.force_torque_valid false; they are not dropped. An arm whose rig declares the sensor but recorded no readings has the columns, all zeros and invalid", "rate_hz_nominal": "about 100 (the real-time report); 5 when the follower fell back to the rich report", "measured_median_hz": { "right": 99.5 }, "measured_per_episode": "sidecar conversion.streams.force_torque..rate_hz; conversion.force_torque_valid_ratio", "hold_max_age_s": 0.2 }, "target_pose": { "feature": "paddy.target_pose.", "pose": "the last TCP target the teleop controller published (/paddy/_arm/target_pose): absolute in the arm base frame (action.frames, TCP action.tcp_frames), m and quaternion xyzw, sign-continuous per episode, held with action_hold_s; what action holds in a tcp dataset. Zeros for an arm driven by joint targets alone (a GELLO in joint mode)" } }, "timing": { "camera_sync_worst_median_ms": 12.499, "command_lag_ms": { "left": 110.35, "right": 122.78 }, "action_latency_ms": { "left": { "arm": 110.35, "gripper": 112.2 }, "right": { "arm": 122.78, "gripper": 187.7 } }, "action_latency_per_episode": "meta/paddy/episodes/{trial_key}.json conversion.action_latency_ms..arm / .gripper" } }, { "path": "policy/datasets/cablesort_2026-11-13_parts_showcase_v2/2026-11-14_02-10-30_trial_005", "trials": [ "trial_005" ], "episode_keys": [ "e4807a6a-5d1f-4626-9e8a-8358e7444cb6" ], "sessions": [ "session_2026-11-14_02-10-30" ], "episodes": 1, "frames": 890, "control_rate_hz": { "left": 67.66971056113066, "right": 68.78276127847292 }, "conventions": { "version": 1, "action": { "feature": "action", "space": "joints", "meaning": "per arm the joint command, zero-order held, then the gripper command: the quantity a policy that drives the arms through target_joints outputs. The TCP target the teleop controller published is paddy.target_pose.; the pose the follower sent is paddy.command_sent.", "joints": "absolute positions of the URDF joints joint1..jointN, base to wrist, in observation.state's order; not a delta, not a TCP pose", "joint_unit": "rad", "joints_source": { "left": "command_sent", "right": "command_sent" }, "joints_sources": { "command_sent": "the joint angles the follower sent (command_sent joints_rad: a set_servo_angle call's target or, in servo mode, a set_servo_angle_j setpoint; the xArm7 joint path, GELLO and policy arms), held with action_hold_s; timing.command_methods says which", "target_joints": "a leader arm's recorded joint targets (target_joints, GELLO), held with action_hold_s", "next_state": "the next row's measured joints (the last row its own): the usual stand-in for a position-controlled arm's joint command when none was recorded (a mode-7 arm, older bags)" }, "joints_source_per_episode": "conversion.action_source. of meta/paddy/episodes/.json", "before_first_command": "the row's measured joints (observation.state), paddy.action_valid false: the arm stays where it is until its first command, as the gripper command holds the measured opening", "layout_per_arm": { "left": [ "joint1", "joint2", "joint3", "joint4", "joint5", "joint6", "joint7", "gripper_cmd" ], "right": [ "joint1", "joint2", "joint3", "joint4", "joint5", "joint6", "joint7", "gripper_cmd" ] }, "frames": { "left": "Left_link_base", "right": "Right_link_base" }, "tcp_frames": { "left": "Left_link_tcp", "right": "Right_link_tcp" }, "arm_order": [ "left", "right" ], "gripper": { "unit": "fraction, 0 closed .. 1 open", "physical": { "left": { "model": "g2", "source": "metadata", "state_closed": 0.84, "state_open": 0.0, "cmd_closed": -10.0, "cmd_open": 850.0 }, "right": { "model": "g2", "source": "metadata", "state_closed": 0.84, "state_open": 0.0, "cmd_closed": -10.0, "cmd_open": 850.0 } } }, "already_scaled": true, "scaling_note": "action is in robot units after the controller's motion scaling and rate limit; paddy.scaling is the human-input gain in force (informational, not a policy input)" }, "timing": { "row_time": "t_k = t0 + k / fps (t0 = the latest first stamp over the arms' joint states and the cameras' color frames; paused spans removed); timestamp = k / fps; paddy.utc_s and the sidecar's conversion.t_start_s give the absolute time", "rule": "row k holds the last joint command stamped at or before t_k (zero-order hold, no shift): the command issued at t_k; with joints_source next_state the measured joints of row k+1, the position the arm reached next", "stamp_meaning": "command_sent: header.stamp of the command_sent message, when the follower's motion call (set_servo_angle, or set_servo_angle_j in servo mode) returned; target_joints: header.stamp of the leader's target; next_state: the next row's slot time", "execution": "the follower validates each target and commands the arm from it; how depends on the arm's control mode, named per command by command_sent.method (per arm command_methods, per episode the sidecar's conversion.streams.command_sent..methods) and described in execution_by_method. The issue time of each command is header.stamp of the command_sent stream", "command_methods": { "left": [ "set_servo_angle_j" ], "right": [ "set_servo_angle_j" ] }, "execution_by_method": { "set_servo_angle_j": "servo mode (xArm mode 1): the follower takes each accepted target as joints (its own solution of a TCP target, or a leader's joint target), steps a joint setpoint toward the newest one on a fixed tick (rates.control.servo_rate_hz), each joint inside the stream's speed and acceleration caps, and sends every setpoint with set_servo_angle_j; the control box executes the last setpoint it received and plans nothing. command_sent has one entry per setpoint: joints_rad is the setpoint, the pose and target_stamp are those of the target it steps toward" }, "action_hold_s": 0.2, "valid_mask": "paddy.action_valid", "issued_mask": "paddy.action_issued" }, "rates": { "video_fps": 30, "control": { "teleop_target_hz_nominal": "50 with the target buffer on, else the Quest sample rate (about 50 to 72)", "target_buffer_ms": { "left": 0.0, "right": 0.0 }, "servo_rate_hz": { "left": 200.0, "right": 200.0 }, "measured_median_hz": { "action_joints.left": 199.45332635883778, "action_joints.right": 222.36793553175698, "command_sent.left": 199.45332635883778, "command_sent.right": 222.36793553175698, "target.left": 67.66971056113066, "target.right": 68.78276127847292 }, "measured_per_episode": "sidecar conversion.streams.target..rate_hz and conversion.streams.command_sent..rate_hz" }, "observation_state_hz": 100, "eef_pose_hz": 100, "safety_state_hz": 10 }, "observation_state": { "joint_unit": "rad", "joint_order": "per arm the URDF joints joint1..jointN, base to wrist, then the gripper fraction; arms in arm_order", "gripper_unit": "fraction, 0 closed .. 1 open", "gripper_opening_m": "per arm, meta/paddy/info.json grippers..opening_m: the opening between the jaws at fraction 0 and at fraction 1, metres; the fraction is linear in it" }, "eef_pose": { "feature": "observation.eef_pose.", "pose": "measured TCP pose in the arm base frame, m and quaternion xyzw, zero-order held", "mode": "urdf", "source": { "left": "urdf_fk", "right": "urdf_fk" }, "sources": "urdf_fk: the recorded URDF's base_T_tcp at the recorded joint states (100 Hz), the kinematics the wrist cameras' paddy.camera_pose and the teleop targets use; reported: the pose the arm's control box reports (about 50 Hz, its own kinematics, which can differ from the URDF's by several mm: sidecar conversion.kinematics..tcp_vs_urdf_mm); mixed: episodes differ, see each sidecar's conversion.tcp_pose.", "stamps": "urdf_fk: the joint states' stamps. reported: the control box reports the pose ahead of the arm; each sample is restamped by that lead, measured per episode against the URDF's TCP at the recorded joints, else the rig profile's actual_pose_lead_ms: sidecar conversion.eef_timing." }, "eef_state": { "feature": "observation.eef_state", "meaning": "every arm's measured TCP pose (observation.eef_pose., sign-continuous) and gripper opening (observation.state's) in one vector, the measured counterpart of a tcp action: the state of an end-effector policy", "layout_per_arm": [ "x", "y", "z", "qx", "qy", "qz", "qw", "gripper" ], "arm_order": [ "left", "right" ], "gripper": "as observation.state's gripper; 0 for an arm without one" }, "kinematics": { "sidecar": "conversion.kinematics. of meta/paddy/episodes/.json", "tcp_vs_urdf": "tcp_vs_urdf_mm and tcp_vs_urdf_deg (median, max): the URDF's base_T_tcp at the recorded joints against the recorded TCP pose", "dh": "dh: the arm control box's own DH table (metadata.json controller_settings..dh_params: 28 values, four per joint, mm and rad) as identified against the recorded TCP pose: layout (the name of each of the four values per joint, in the table's order), convention (modified = Craig, or standard), tcp_vs_dh_mm and tcp_vs_dh_deg (median, max), samples, fit, and dh_vs_urdf_mm (the table's flange against the URDF's at the same joints); with fit true, forward kinematics of the table read that way is the arm's calibrated kinematics. {note} when the trial has no table" }, "cameras": { "file": "meta/cameras.json", "role": "where the camera is mounted, as the rig profile declares it: overhead (fixed, looking down at the workspace), scene (fixed, another viewpoint) or wrist (moves with an arm; which arm is mount.side); null = not declared", "primary": "the file's top-level primary names the rig's one primary camera (the view to look at first), and that camera's entry has primary: true; null = not declared", "base_T_cam": "4x4 pose of the camera's color optical frame (z forward, x right, y down) in each arm base frame, metres; extrinsics..source says where it came from", "sources": { "tf_static": "the nominal pose: the rig's URDF and the camera mounts in the rig profile (CAD or tape), as published on /tf_static", "tf_snapshot": "the nominal pose from the first /tf snapshot at episode start (valid: episode_start)", "urdf_fk": "the nominal pose of a wrist camera by forward kinematics of the recorded URDF and joint states at t_s, the episode's first slot (valid: episode_start)", "measured": "the rig profile's measured calibration: calibration.frame_T_cam chained to the base through via (identity, tf_static, urdf, urdf_fk at t_s for a wrist camera, or tf_snapshot); the nominal entry is kept as nominal {base_T_cam, source}" }, "measured_vs_nominal": "measured_vs_nominal_mm / _deg: distance of the origins and angle of the rotation between the measured and the nominal base_T_cam at the episode's first slot", "intrinsics": "K (row-major 3x3) and D (plumb_bob k1 k2 p1 p2 k3) describe the raw color image as recorded (not undistorted). intrinsics.source camera_info: the camera's own CameraInfo (a RealSense's factory values; a D435 color stream reports D all zeros). measured: the rig profile's measured intrinsics (method, calibrated_at, error_px, samples, notes, from), the CameraInfo values kept as factory_intrinsics {K, D, distortion_model}. The pose of a fixed camera measured after its intrinsics were is solved with them", "calibration": "per camera: the measured calibration (frame, xyz_mm, rpy_deg URDF fixed-axis, frame_T_cam, method, calibrated_at, error_mm, error_px, samples, notes, from: rig_snapshot | rig_option), null when the rig profile has none. TF on the rig stays nominal", "image_rotation": "image_rotation_deg per camera: 0, or 180 for a camera mounted upside down whose pictures the dataset turns. Turned, the color and depth frames are the recorded ones turned by 180 degrees and every value that describes them is of the turned image: K (cx' = W - 1 - cx, cy' = H - 1 - cy), D (p1 and p2 negated), base_T_cam, mount link_T_cam and tcp_T_cam, calibration.frame_T_cam, the depth K, D and color_T_depth, and paddy.camera_pose.: the camera's own optical frame (frame_id) turned about its z axis, X_T_turned = X_T_cam @ diag(-1, -1, 1, 1). calibration.xyz_mm and rpy_deg stay the rig profile's record of the camera's own optical frame. A policy trained on it needs the live frames turned the same way (paddy's policy server does it from this file)", "per_row_pose": "paddy.camera_pose. of a wrist camera; source measured_fk (forward kinematics to the calibration frame times its measured pose), urdf_fk or eef; time row (at the joint sample observation.state holds, so it matches observation.state through the URDF; the image is paddy.frame_age. older) or frame (at the shown frame's stamp)" }, "subtasks": { "feature": "paddy.subtask_index", "meaning": "the operator-marked subtask open at the row (metadata.json subtasks): its index, -1 where none is open. Spans are half-open [start, end) in header.stamp POSIX seconds; paused spans are out of the grid, so a subtask with a pause inside has no rows for the paused time", "names": "the episode sidecar's conversion.subtasks: per index the plain-language instruction, the outcome (success, failure or unknown), start_s, end_s (null: still open at the end of the recording) and the rows [start_row, end_row) it covers (null when none)", "not_a_policy_input": true }, "command_sent": { "feature": "paddy.command_sent.", "pose": "the pose the follower sent to the arm (its mm / Euler call converted back to m / quaternion xyzw in the base frame; for an xArm5 the projected pose actually sent), zero-order held with action_hold_s; the call's speed and acceleration are in the sidecar's conversion.motion_profile.", "before_first_command": "the measured TCP pose (observation.eef_pose.), paddy.action_issued false: the arm stays where it is until its first command. The TCP target (a tcp dataset's action, paddy.target_pose. in a joints dataset) holds the measured TCP pose before the first target too, paddy.action_valid false. Zeros only for an arm whose stream, or TCP pose, was never recorded" }, "velocity": { "feature": "observation.velocity", "status": "present", "unit": "rad/s", "source": "sdk", "meaning": "joint velocities the xArm SDK reports (get_joint_states), as the follower published them with each joint-state sample (100 Hz), zero-order held like observation.state", "order": "per arm joint1..jointN (URDF order, base to wrist), as observation.state; no gripper" }, "effort": { "feature": "observation.effort", "status": "present", "unit": "N m", "source": "sdk_estimate", "validated": false, "meaning": "the xArm SDK's joint torque estimate (get_joint_states) from motor current and the arm's own dynamic model; the arms have no joint torque sensors. Sign and magnitude not yet checked on the rig", "order": "per arm joint1..jointN (URDF order, base to wrist), as observation.state; no gripper", "validation": "static torques at four poses against the URDF masses' gravity torques, then the gripper pressed on a kitchen scale" }, "joint_state_status": "present: the column is written; absent: a bag's joint states carry none, so the dataset leaves it out (never zeros; each sidecar's conversion.joint_velocity / joint_effort names the episodes); off: turned off (--no-joint-velocity, --no-joint-effort)", "rot6d": "first_two_columns_column_major", "rot6d_features": { "layout": [ "r00", "r10", "r20", "r01", "r11", "r21" ], "meaning": "the first two columns of the rotation matrix R (base_T_tcp's rotation), column-major: [r00, r10, r20, r01, r11, r21] (Zhou et al. 2019; GR00T XYZ_ROT6D, OpenNeoData). Continuous by construction; recover R by Gram-Schmidt on the two columns and their cross product. Zeros where the quaternion is zero (a stream that was never recorded)", "observation": "observation.eef_rot6d.", "target": "paddy.target_rot6d.", "not_a_policy_output": "the target's 6D rotation is outside the action prefix" }, "force_torque": { "feature": "observation.force_torque.", "raw_feature": "paddy.force_torque_raw.", "valid_mask": "paddy.force_torque_valid", "arm_order": [ "right" ], "layout": [ "fx", "fy", "fz", "tx", "ty", "tz" ], "units": { "force": "N", "torque": "N m" }, "frames": { "right": "Right_link_eef" }, "frame_note": "the flange (URDF link_eef), which the six-axis sensor is mounted on. Checked on the rig (2026-11-07): a press along each flange axis reads on that axis with the sign of the force applied to the tool (a push toward flange +x reads +fx); a load fit put the sensor axes within 1.4 deg of the flange's", "reading": "observation.force_torque.: the control box's reading after its load and offset compensation (/paddy/_arm/ft_wrench, SDK ft_ext_force): near zero, up to a small offset, with nothing touching the tool when the payload set on the control box matches the tool. paddy.force_torque_raw.: the unprocessed reading (/paddy/_arm/ft_wrench_raw, SDK ft_raw_force), so the compensation can be redone", "compensation": "the payload the control box compensated for is in each episode sidecar's metadata.controller_settings..tcp_load_kg and tcp_load_cog_mm, read at the trial's start", "timing": "row k holds the last reading stamped at or before t_k (zero-order hold, fresh within hold_max_age_s); the stamp is when the follower's host read the control box's real-time report packet", "missing": "rows without a fresh reading hold zeros and paddy.force_torque_valid false; they are not dropped. An arm whose rig declares the sensor but recorded no readings has the columns, all zeros and invalid", "rate_hz_nominal": "about 100 (the real-time report); 5 when the follower fell back to the rich report", "measured_median_hz": { "right": 99.5 }, "measured_per_episode": "sidecar conversion.streams.force_torque..rate_hz; conversion.force_torque_valid_ratio", "hold_max_age_s": 0.2 }, "target_pose": { "feature": "paddy.target_pose.", "pose": "the last TCP target the teleop controller published (/paddy/_arm/target_pose): absolute in the arm base frame (action.frames, TCP action.tcp_frames), m and quaternion xyzw, sign-continuous per episode, held with action_hold_s; what action holds in a tcp dataset. Zeros for an arm driven by joint targets alone (a GELLO in joint mode)" } }, "timing": { "camera_sync_worst_median_ms": 14.098, "command_lag_ms": { "left": 121.05, "right": 115.29 }, "action_latency_ms": { "left": { "arm": 121.05, "gripper": 127.5 }, "right": { "arm": 115.29, "gripper": 106.9 } }, "action_latency_per_episode": "meta/paddy/episodes/{trial_key}.json conversion.action_latency_ms..arm / .gripper" } } ] } }