LeeHakHo/pickplace_real101

Pick-and-place — 101 real Franka demonstrations

Real-robot pick-and-place for the ID90 / OOD auxiliary-pose study, recorded on the same earlgrey rig as LeeHakHo/square_real100 and built with the same pipeline. A Franka Panda picks up two objects in turn from the large compartment of a wooden tray and places each into one of the four small compartments next to it:

1. a matte red cylinder (5.3 cm diameter x 10.8 cm tall), always first, always into the same compartment;
2. a green cube (5 cm), always second, always into the compartment beside it.

Every episode is a successful placement of both objects.

Demonstrations	101
Samples	64,592
Recording rate	nominally 15 Hz; measured 14.4 Hz (69.4 ms per step)
Episode length	399 – 996 frames (median 614)
Size	1.91 GB
Format	robomimic HDF5

Layout

data/demo_<i>/
  obs/agentview_image            (T,84,84,3) uint8   Intel RealSense D435, static, scene
  obs/robot0_eye_in_hand_image   (T,84,84,3) uint8   ZED Mini, wrist-mounted
  obs/robot0_eef_pos             (T,3)   float64     panda_hand_tcp, robot base frame
  obs/robot0_eef_quat            (T,4)   float64     xyzw
  obs/robot0_gripper_qpos        (T,2)   float64     [+f, -f], robosuite convention
  obs/object                     (T,14)  float64     block 0 = cylinder, block 1 = cube,
                                                     each [pos(3) | quat(4) xyzw]
  obs/aux_valid                  (T,2)   uint8       column b gates block b -- see below
  actions                        (T,7)   float64     [dpos(3) | drot axis-angle(3) | grasp]
  rewards, dones                 (T,)                sparse: 1 on the final frame
  joint_qpos                     (T,7)   float64     arm joints, provenance
mask/train  (91 demos)   mask/valid  (10 demos)

Each demo group carries source_session, source_traj and object_slugs (["red_cylinder_object", "green_cube"], the block order). demo_<i> is traj_<i> of the raw session. Images were captured at 1280x720, centre-cropped to square and resized to 84x84. The crop contains the whole tray: object centroids reach the crop edge in 0.01% of frames.

For an auxiliary pose head: algo.aux_pose.obj_blocks = [0, 1], and supervise no rotation for either object (ROTATION_BLOCKS = [], no --aux_rotation). The cylinder is a body of revolution, so its yaw is not a function of the image. That is the same failure the simulated study handles for coffee_pod, and the cube's yaw is dropped along with it. The quaternions are kept in the file but FoundationPose's yaw for the cylinder is arbitrary.

obs/aux_valid matters much more here than on Square

obs/aux_valid[:, b] is 1 where block b of obs/object can be trusted. Multiply the per-frame auxiliary loss for each object by its own column. It is one column per object rather than one ANDed mask, because the cube is routinely mis-tracked while the cylinder is fine, and ANDing would throw away good cylinder labels. Check that your training code indexes a (T,2) mask per block: Square's is (T,).

	frames valid	per demo
cylinder	99.1%	median 100%, min 81.7%
cube	70.2%	median 78.8%, min 2.4%; 23 demos below 60%

Where the cube's labels survive, by task phase:

phase (share of frames)	cylinder valid	cube valid
start → cylinder grasp (26%)	99%	83%
cylinder being carried (23%)	97%	81%
between the two grasps (28%)	100%	81%
cube being carried (19%)	100%	21%
after the cube is released (3%)	100%	85%

Why the cube fails

obs/object is FoundationPose tracking RGB-D, not ground truth. The cylinder tracks well. While it is carried, it stays 7.7 mm (median) from where rigid attachment to the gripper puts it. The 5 cm cube does not, in two ways:

• While it is carried (98 of 101 demos): the pose stays on the cube in the image but slides along the camera ray to roughly table height. The median error against rigid attachment is 75 mm, up to 24 cm. No single step jumps, so a jump test cannot see it.
• While it waits and the arm carries the cylinder over it, the pose drifts (median 41 mm, max 181 mm), then usually recovers.

Four cylinder carries fail the same way (demo_5, 20, 83, 94) and are masked.

How the mask was derived

From physical facts about the task, with no hand labelling. The exact procedure is aux_validity() in panda_utils/scripts/r2st_to_robomimic.py, run with --carry-tol-m 0.02 --pre-grasp-ref-frames 10 --carry-anchor-tol-m 0.05.

1. No teleporting. An object cannot move 3 cm between consecutive frames.
2. Static before its own grasp. Within that window, a frame more than 2 cm from the object's frame-0 pose is invalid. Frame 0 is where FoundationPose registers, with the arm parked away. Anchoring there, and not at the window median, matters: in 5 demos the drifted pose held most of the window.
3. Static after its own release. A frame more than 2 cm from that window's median is invalid.
4. Rigid with the gripper while grasped. The object's position expressed in the end-effector frame may not move more than 2 cm from its value 5–10 frames after the fingers close. By then the fingers have finished closing but the arm has not yet lifted. An earlier reference would misread the fingers centring the object as error. If the object at that moment is more than 5 cm (xy) from its frame-0 pose, the whole carry is invalid, because the reference itself is wrong. The 5 cm allows for the approach bumping the object: the well-tracked cylinder moved up to 48 mm before its grasp.

Frames within ±2 of a flagged frame are also invalid. Each object's windows come from its own grasp. The grasp intervals are found from the gripper signal and assigned to objects by which one moved during each. That assignment is read from the data, although here it is always cylinder first.

Checks 2 (frame-0 anchor) and 4 are not applied to square_real100. Measured on Square, the nut stays within 7.5 mm (median) of the gripper while carried. Its larger deviations come from real motion: the fingers centring the handle, and a ~30° swing about the pinch point.

Two recording sessions, one camera pose each

	demos (by source_traj)	recorded
A	0–3, 5–28, 31–43, 45, 46 (43)	2026-09-09 23:50 – 09-10 01:48
B	4, 29, 30, 44, 47–100 (58)	2026-09-10 20:08 – 22:36

The trajectory numbers were reassigned after bad recordings were removed, so they are not chronological. The D435 moved slightly between the sessions. Background SIFT on frame 0, lifted to 3D with depth and solved by PnP, puts the motion at 0.50° and 6 mm. Within a session the same measurement is 0.09° / 2.9 mm. At the tray centre that moves the object label by **4 mm**, similar in size to FoundationPose's own static jitter, so both sessions use one extrinsics. For the 84 px policy input the shift is ~0.3 px, well inside random-crop augmentation. Label quality does not differ by session (cube valid A 66.9%, B 72.6%).

Excluded before conversion: 8 raw recordings, with a frozen wrist camera, a truncated file, the gripper starting closed, or not starting from the reset pose.

Task layout

Object positions in the robot base frame (metres), from the tracked poses:

	x	y
cylinder start	0.41 – 0.63	−0.23 – −0.10
cube start	0.40 – 0.63	−0.23 – −0.10
cylinder placed	0.51 – 0.63	+0.17 – +0.25
cube placed	0.49 – 0.64	−0.02 – +0.11

Start positions lie on a grid, about x ∈ {0.43, 0.50, 0.56, 0.63} by y ∈ {−0.21, −0.16, −0.10}, with 53 distinct cylinder/cube cell combinations across the 101 demos and at least 8 cm between the objects. A position between grid points is interpolation, not in-distribution coverage. Keep that in mind when defining an OOD ladder's in-distribution window.

Two things that will bite you (same as Square)

There is no simulator behind this data. data.attrs["env_args"] (PickPlaceReal) records provenance only. Set experiment.rollout.enabled = false.

Actions are reconstructed, not recorded. The logger never stores the teleoperation command, so:

dpos  = eef_pos[t+1] - eef_pos[t]              robot base frame
drot  = axis-angle of R[t].T @ R[t+1]          as OSC_POSE expects
grasp = +1 while the fingers are closed, -1 while open

These are per-step deltas at the measured 14.4 Hz. Run the controller at that rate, or account for the 4% difference. The grasp label flips when the fingers pass the midpoint of their travel, 2–3 frames after they start moving. Measured against the demos, operators lift 11 frames (median) after the flip while the fingers close in 5, so a policy reproducing the timing still has a ~7-frame margin.

Calibration and accuracy

Same ChArUco hand-eye calibration as square_real100 (2.58 mm RMS over five touched corners). Known residual error sources:

• D435 depth reads ~14 mm farther than RGB geometry on this table, so tracked positions carry roughly that bias along the camera ray.
• Meshes use measured dimensions (cylinder 5.3 x 10.8 cm without taper, cube 5 cm). scale_glb_to_pointcloud rescales them per demo onto the observed point cloud.
• The camera moved ~4 mm at the tray between sessions (above), and by 0.43° / 8.4 mm between the Square sessions and this collection. A constant offset like this is absorbed by an auxiliary head's output bias.