bkuster's picture
Publish combined_v0: four pinned datasets, 1975 episodes and 146240 frames
f9184eb verified
|
Raw History Blame Contribute Delete
10.4 kB
metadata
license: apache-2.0
task_categories:
  - robotics
tags:
  - lerobot
  - oracle
  - franka
configs:
  - config_name: default
    data_files: data/*/*.parquet

Elementary Franka oracle demonstrations

1,069 episodes / 32,070 frames / 98 instructions at 15 FPS.

Group Instructions Episodes per instruction Episodes
Original movements and stiffness variants 48 15 720
Standalone open/close gripper 2 20 40
Movements with the gripper closed 48 6 or 7 309

The closed-gripper extension is partially collected: 309 of the planned 480 episodes, with 171 remaining. Its task distribution is not exactly balanced. This snapshot is ready for training on the collected data; exact counts are below.

Observation and action contract

LeRobot v3.0; Franka Research 3; RGB 640×480 under observation.images.image. observation.state has 31 floats: measured Cartesian pose (XYZ + rotation-6D), commanded pose (XYZ + rotation-6D), measured base-frame Cartesian velocity (6), measured tool-frame wrench (6), and episode elapsed seconds (1). Measured gripper width is not in the state.

action has 11 floats: six tool-frame motion channels, normalized gripper closure, three stiffness factors, and done. Despite their cart_pos_*_tool_commanded_rel names, translation channels are velocity-like commands in m/s, integrated at dt=1/15 s; angular channels are rad/s. Gripper 0=open, 1=closed. XY/Z stiffness factors map to 100–1500 N/m; rotation maps to 10–30 Nm/rad. Done is zero for these elementary tasks. Exact feature names/order are in meta/info.json.

Recording behavior

Original movement instructions command open fingers. Closed-gripper variants start with a verified 60 mm open gap, prepared before recording. Arm movement and asynchronous closure start together on the first recorded action and continue for two seconds. Empty-jaw closing uses positional Move to zero; no object-contact grasp is required. Initial/final gaps are in closed_gripper_setup.jsonl and closed_gripper_endpoints.jsonl. Standalone open tasks start closed; standalone close tasks start open.

Movement speed is 0.03 m/s or 0.15 rad/s after the 1.5 oracle multiplier. Starts use uniform independent offsets ±5 cm XYZ, ±2° roll, no pitch offset, ±5° yaw around the saved reference. Cartesian resets take six seconds plus a 0.5-second settling wait. Resets and initial opening are not recorded. These are elementary demonstrations, not pickup episodes.

Stiffness settings changed during collection; older episodes retain their actual historical labels. Later recordings cap rotation at 25 Nm/rad and scale physical XY/Z gains by 0.9, respecting the 100 N/m minimum: ordinary gains 900 N/m, maximum 1350 N/m; low-rotation tasks can use 10 Nm/rad. Task text alone does not uniquely determine every historical gain. Applied stiffness factors are in each action, with batch arguments/manifests in source_metadata/.

Validation and provenance

Every retained video frame was decoded; data indices, metadata shard references, video ranges, numeric finiteness, gripper endpoints, and dataset-reader boundaries were checked. No identical complete two-second video sequences were found in the final snapshot. Known camera-disconnection episodes were removed. See publication_validation.json, camera_filter_manifest.json, and the historical cleanup manifests.

Filtering IDs refer to the historical snapshot identified in each manifest; later valid appends may reuse those IDs. Historical manifests and the original recording request can show older counts or targets. meta/info.json and publication_validation.json describe the published snapshot. Backups, rejected recordings, caches, credentials and model weights are excluded.

Structural validation and endpoint checks are not independent physical-success annotations, and do not rule out every transient stale camera frame. Camera reads use the latest cached image without blocking. Joint/nullspace motion is not fully represented by the Cartesian state.

Training

Dataset ID: bkuster/oracle_basic_15pertask_20261109_094952. Use the 31-state/11-action contract and RGB key above. Metadata currently puts all episodes in train; make any validation split by episode rather than individual frames. meta/stats.json covers all retained data, so recompute training-only normalization after splitting or merging. Pin the publication commit for reproducibility.

Exact task counts

Task instruction Episodes
Move in gripper positive x 15
Move in gripper negative x 15
Move in gripper positive y 15
Move in gripper negative y 15
Move in gripper positive z 15
Move in gripper negative z 15
Rotate around gripper positive x 15
Rotate around gripper negative x 15
Rotate around gripper positive y 15
Rotate around gripper negative y 15
Rotate around gripper positive z 15
Rotate around gripper negative z 15
Move in base positive x 15
Move in base negative x 15
Move in base positive y 15
Move in base negative y 15
Move in base positive z 15
Move in base negative z 15
Rotate around base positive x 15
Rotate around base negative x 15
Rotate around base positive y 15
Rotate around base negative y 15
Rotate around base positive z 15
Rotate around base negative z 15
Open gripper 20
Close gripper 20
Move in base positive x with low XY stiffness 15
Move in base positive x with maximum stiffness 15
Move in base negative x with low XY stiffness 15
Move in base negative x with maximum stiffness 15
Move in base positive y with low XY stiffness 15
Move in base positive y with maximum stiffness 15
Move in base negative y with low XY stiffness 15
Move in base negative y with maximum stiffness 15
Move in base positive z with low Z stiffness 15
Move in base positive z with maximum stiffness 15
Move in base negative z with low Z stiffness 15
Move in base negative z with maximum stiffness 15
Rotate around base positive x with low rotational stiffness 15
Rotate around base positive x with maximum stiffness 15
Rotate around base negative x with low rotational stiffness 15
Rotate around base negative x with maximum stiffness 15
Rotate around base positive y with low rotational stiffness 15
Rotate around base positive y with maximum stiffness 15
Rotate around base negative y with low rotational stiffness 15
Rotate around base negative y with maximum stiffness 15
Rotate around base positive z with low rotational stiffness 15
Rotate around base positive z with maximum stiffness 15
Rotate around base negative z with low rotational stiffness 15
Rotate around base negative z with maximum stiffness 15
Move in gripper positive x with the gripper closed 7
Move in gripper negative x with the gripper closed 7
Move in gripper positive y with the gripper closed 7
Move in gripper negative y with the gripper closed 7
Move in gripper positive z with the gripper closed 7
Move in gripper negative z with the gripper closed 7
Rotate around gripper positive x with the gripper closed 7
Rotate around gripper negative x with the gripper closed 7
Rotate around gripper positive y with the gripper closed 7
Rotate around gripper negative y with the gripper closed 7
Rotate around gripper positive z with the gripper closed 7
Rotate around gripper negative z with the gripper closed 7
Move in base positive x with the gripper closed 7
Move in base negative x with the gripper closed 7
Move in base positive y with the gripper closed 7
Move in base negative y with the gripper closed 7
Move in base positive z with the gripper closed 7
Move in base negative z with the gripper closed 7
Rotate around base positive x with the gripper closed 7
Rotate around base negative x with the gripper closed 7
Rotate around base positive y with the gripper closed 7
Rotate around base negative y with the gripper closed 6
Rotate around base positive z with the gripper closed 6
Rotate around base negative z with the gripper closed 6
Move in base positive x with low XY stiffness with the gripper closed 6
Move in base positive x with maximum stiffness with the gripper closed 6
Move in base negative x with low XY stiffness with the gripper closed 6
Move in base negative x with maximum stiffness with the gripper closed 6
Move in base positive y with low XY stiffness with the gripper closed 6
Move in base positive y with maximum stiffness with the gripper closed 6
Move in base negative y with low XY stiffness with the gripper closed 6
Move in base negative y with maximum stiffness with the gripper closed 6
Move in base positive z with low Z stiffness with the gripper closed 6
Move in base positive z with maximum stiffness with the gripper closed 6
Move in base negative z with low Z stiffness with the gripper closed 6
Move in base negative z with maximum stiffness with the gripper closed 6
Rotate around base positive x with low rotational stiffness with the gripper closed 6
Rotate around base positive x with maximum stiffness with the gripper closed 6
Rotate around base negative x with low rotational stiffness with the gripper closed 6
Rotate around base negative x with maximum stiffness with the gripper closed 6
Rotate around base positive y with low rotational stiffness with the gripper closed 6
Rotate around base positive y with maximum stiffness with the gripper closed 6
Rotate around base negative y with low rotational stiffness with the gripper closed 6
Rotate around base negative y with maximum stiffness with the gripper closed 6
Rotate around base positive z with low rotational stiffness with the gripper closed 6
Rotate around base positive z with maximum stiffness with the gripper closed 6
Rotate around base negative z with low rotational stiffness with the gripper closed 6
Rotate around base negative z with maximum stiffness with the gripper closed 6