ajaysri/pi07_cable_three_grasp_place_v1

Three-holder cable routing: grasp, place, and route

LeRobot v3 demonstrations with separate numbered grasp, placement, and routing subtasks. Grasp subtasks use a clean overhead goal image with no vector overlay. Placement and routing retain the recorded target-vector overlay.

Property	Value
Original recordings	140
Subtask episodes	1,260
Frames	335,897
Sampling rate	100 Hz
Robot	ARX bimanual
State / action dimensions	14 / 14
Camera image resolution	448 × 448 RGB
Split	Train only

Each recording contributes nine episodes, in this order:

1. grasp holder 1
2. place holder 1
3. route rope through holder 1
4. grasp holder 2
5. place holder 2
6. route rope through holder 2
7. grasp holder 3
8. place holder 3
9. route rope through holder 3

These exact strings are stored in prompt, metadata.phase_label, and the task vocabulary. Task indices are 0–8.

Boundaries and keyframes

The selected grasp frame is included in the grasp subtask. Placement begins on the next frame. Original placement-end and routing boundaries are retained, and every recorded frame appears exactly once.

Recording episode_00055.hdf5, holder 3 uses manually selected frame 1904: grasp ends at 1904 inclusive, and placement begins at 1905. The gripper is unspecified because this selection was made from the overhead view.

The remaining 419 grasp boundaries come from sustained gripper-closure detection. These are heuristic candidates; this release does not claim that all boundaries were manually reviewed or that overhead visibility was automatically verified. The source's optional pick phases are combined with placement before applying the selected grasp boundary consistently across recordings.

annotations/grasp_selections.json records the selected frames, optional arms, and selection methods. annotations/subtasks.json records every subtask's source-frame interval, inclusive end keyframe, and parent recording index. metadata.keyframe_source_index repeats the subtask's end keyframe on each frame row; it is annotation metadata, not a current policy observation.

Camera and goal streams

LeRobot video feature	Content
observation.images.video_overhead_raw	Clean current overhead image
observation.images.video_left	Current left wrist image
observation.images.video_right	Current right wrist image
observation.images.video_overhead_vector	Clean overhead during grasp; vector-annotated overhead during placement and routing

The fourth feature retains the parent's name for compatibility. Its content depends on the episode. Use the video file index and timestamp offsets in episode metadata. During grasp, its file index selects a byte-identical copy of the clean overhead stream. During placement and routing, it selects the original vector stream. Inferring a goal-video path from the recording index alone is incorrect.

metadata.goal_vector_visible is 0 for grasp and 1 for placement/routing. Vector coordinates remain as recorded placement-target metadata for all rows; they are not grasp-point annotations. Native images were resized from 1280 × 800 to 448 × 280 and vertically padded to 448 × 448. Placement/routing arrows have the same 28-pixel length at this resolution. This revision copies encoded video bytes without re-encoding.

All video streams store images at their recorded timestamps. Any past-goal sampling is performed by the training loader and must stay within the new subtask. Overhead keyframe annotations can be used to prepare high-level-policy supervision; this repository is the LeRobot dataset, not a model-specific HLP training corpus or trained checkpoint.

Actions and normalization

Robot state, velocity, action, and original source-phase metadata are preserved. Local frame_index and timestamp restart at each new subtask; source frame indices and action timestamps retain the original recording coordinates.

Action chunks must stop at the new subtask boundary. metadata.valid_future_action_steps counts valid future actions up to a 100-step horizon, including the current action. The normalization statistics were recomputed for uniform frame sampling with 100-step chunks padded by repeating the subtask's final action. A training consumer must honor these boundaries and the new nine-label vocabulary; the parent's frozen training contract does not apply unchanged.

All source recordings are in the training split. No held-out evaluation or novelty claim is made.

Download and inspect

from huggingface_hub import snapshot_download

root = snapshot_download(
    repo_id="ajaysri/pi07_cable_three_grasp_place_v1",
    repo_type="dataset",
)

Pin revision= to a Hub commit for reproducible use. A LeRobot v3-compatible reader must use meta/episodes/chunk-000/file-000.parquet for the shared data-file and video-file references and their offsets. The Hugging Face table viewer displays per-frame records; synchronized videos are stored separately under videos/.

This dataset derives from the original vector-goal dataset. contract.json contains input revisions, the annotation derivation, representation rules, and SHA-256 hashes of dataset artifacts. The original dataset is unchanged.

Dataset content revision: 9cc094d403628104152dac9f20e43688b33c2a436aab7921e1df8923776c6004.