ajaysri/route_plus_steer_place_subtasks_dual_overhead_pi05_224_plus_dagger

224px Route + Steering Placement + DAgger with Dual Overhead Views

This is a LeRobot v2.1 training dataset at 100 Hz. Every output episode contains one constant prompt. Base route boundaries are inferred from maximal contiguous runs of metadata.route.active_subgoal; partial source episodes are kept as the suffix subtasks they contain.

The base data contains 495 route subtask episodes (128742 frames) and 202 steering/placement episodes (50099 frames). All 66 route DAgger clips and all of their frames are appended 12 times for 792 DAgger episode instances and 26964 weighted DAgger frames. The combined dataset has 1489 episodes and 205805 frames; DAgger contributes 13.1% of frames.

Camera streams

• observation.images.video_overhead_vector: overhead image with the fixed yellow-dot/red-arrow goal.
• observation.images.video_overhead_raw: the same overhead frame without the synthetic vector.
• observation.images.video_left and observation.images.video_right: raw side views.

For base route and steering episodes, all four streams are encoded at 224x224 from the cameras' raw 1280x800 JPEG frames. Each raw frame is aspect-preservingly downsampled to 224x140 with Lanczos filtering, then padded with 42 black pixels above and below. No frame is upsampled, cropped, or stretched. The vector stream differs from the raw overhead stream only by the synthetic goal glyph.

The DAgger source is deliberately different: its four feeds were already letterboxed at 224x224 and are passed through without spatial resampling. The stored camera feeds map directly to left, right, overhead raw, and overhead vector. Every DAgger clip remains whole, including autonomous prefix or suffix frames, and keeps one constant route/place prompt.

For DAgger, the training action column is copied from HDF5 /action; /policy_target_action is not used. The declared training timestamp is always frame_index / 100. The independently measured /action_timestamps are normalized to start at zero and retained in source.action_time_s, so source timing remains available without changing the declared 100 Hz training cadence.

Each unique DAgger video is encoded once and materialized as twelve hard-linked local episode copies. Copy identity and provenance are recorded in metadata.dagger.is_dagger, metadata.dagger.copy_index, metadata.dagger.demo_id, and metadata.dagger.rollout_id. Base episodes use false and -1 sentinels. DAgger connector and subgoal identifiers that were not recorded also use -1.

The raw/vector overhead streams are frame-aligned with each other, the robot state, and the action. Route and steering use their original visual styles, recorded in metadata.vector.style_name and the per-frame size fields. The metadata.vector.source_xy_224 field remains in the historical 224-space letterbox coordinate system. Rendered glyph sizes are scaled to the historical 224px style and reported in output pixels.

Tasks

• place the connector on the selected vector
• route the cable through the selected connector
• move the connector to the yellow dot and match the red arrow orientation

No intervention-only filtering, temporal interpolation, long-cable data, or steering upsampling is applied. DAgger's only intentional weighting is the twelve total copies per unique clip.