Route Subtasks + Balanced DAgger Interventions, 448px
This is a new LeRobot v2.1 dataset derived from ajaysri/route_subtasks_dual_overhead_pi05_448 and a subsequent DAgger collection. The original dataset is not modified.
The base contributes 495 episodes and 128,742 frames. Only
frames recorded while the human collector was actively intervening are added;
autonomous policy-control frames and policy_target_action are excluded from
the training targets. The DAgger action column comes from HDF5 /action.
Data split and weighting
This release has a single train split with 713 episode
instances and 164,019 frames. It has no validation or test
split.
| Source | Unique episodes/segments | Copies per item | Dataset episodes | Dataset frames |
|---|---|---|---|---|
| Original route dataset | 495 | 1 | 495 | 128,742 |
| Placement interventions | 34 | 2 | 68 | 12,966 |
| Routing interventions | 50 | 3 | 150 | 22,311 |
| Total | 713 | 164,019 |
| Task | Original frames | Weighted intervention frames | Total frames | Intervention share |
|---|---|---|---|---|
| Placement | 46,542 | 12,966 | 59,508 | 21.8% |
| Routing | 82,200 | 22,311 | 104,511 | 21.3% |
The intervention collection contains 84 unique human segments and 13,920 unique human frames. Placement segments are represented 2x and routing segments 3x. This makes weighted intervention data 21.5% of training frames while keeping the within-subtask intervention shares close: 21.8% for placement and 21.3% for routing. Copies change sampling exposure, not the amount of unique intervention data.
Camera streams
Every episode contains four synchronized 448x448 H.264 streams:
The preview shows the raw overhead observation on the left and the aligned yellow-dot/red-arrow steering view on the right.
observation.images.video_overhead_vector: overhead observation with the yellow-dot/red-arrow steering goal.observation.images.video_overhead_raw: the same overhead observation without the synthetic steering goal.observation.images.video_left: raw left camera.observation.images.video_right: raw right camera.
Native 1280x800 frames are aspect-preservingly downsampled to 448x280 and letterboxed with 84 black pixels above and below. The vector view is rendered from the recorded per-frame goal center and direction. No frame is cropped, stretched, or upsampled.
Episode boundaries and provenance
Each DAgger intervention segment is a separate episode, so action chunks never
cross from human intervention into policy control. metadata.dagger.is_dagger
is true for appended intervention episodes; base episodes retain false.
metadata.dagger.copy_index, metadata.dagger.demo_id,
metadata.dagger.rollout_id, the source episode/frame fields, and the source
start/end fields preserve provenance and identify weighted copies.
The declared training cadence is 100 Hz and timestamp = frame_index / 100.
source.action_time_s retains the recorded action timing relative to the start
of its original DAgger rollout.
Tasks
place the connector on the selected vectorroute the cable through the selected connector
- Downloads last month
- 148