ajaysri/route_red_yellow_vector_subtasks_pi05
Route Red-Yellow Vector Subtasks for pi0.5 This is a LeRobot v2.1 transformation of DistantSky/route at commit aced1e96f6b8bf98ffaa0754407636e82442b084. The dataset contains 212 real-robot episodes, 135699 frames, three cameras, and 14-dimensional actions at 100 Hz. Conditioning Only observation.images.video_overhead is modified. The left and right videos are byte-identical to the source dataset. The overhead image receives one fixed selected-connector pose glyph… See the full description on the dataset page: https://huggingface.co/datasets/ajaysri/route_red_yellow_vector_subtasks_pi05.
Route Red-Yellow Vector Subtasks for pi0.5
This is a LeRobot v2.1 transformation of DistantSky/route at commit aced1e96f6b8bf98ffaa0754407636e82442b084.
The dataset contains 212 real-robot episodes, 135699 frames, three cameras, and 14-dimensional actions at 100 Hz.
Conditioning
Only observation.images.video_overhead is modified. The left and right videos are byte-identical to the source dataset.
The overhead image receives one fixed selected-connector pose glyph per frame:
- yellow position dot: RGB (255, 230, 0), radius 2.4 px
- red direction arrow: RGB (255, 24, 28), length 14.0 px, width 2.0 px, head 4.0 px
- 4x antialiasing followed by Lanczos downsampling
- no text rendered into image pixels
The frame-level prompt and task_index fields are:
- task 0:
place the connector on the selected vector - task 1:
route the cable through the selected connector
Phases 0 and 1 select connector 1; phases 2 and 3 select connector 2. The vector remains visible during both placement and routing, then switches when connector 2 placement begins. Partial source episodes retain the vector and prompt for their recorded active phase.
Audit fields
The Parquet files retain the original timing metadata and add:
metadata.route.selected_connector_indexmetadata.route.overlay_source_xy_224metadata.route.overlay_direction_xy_224metadata.route.overlay_theta_image_radmetadata.route.overlay_visiblemetadata.route.phase_labelmetadata.route.text_subtask- overlay style and size constants
The target image point is taken directly from the collected router label. Direction is projected from the board-coordinate unit vector through an episode-specific planar homography fitted from the three obstacle and two router correspondences.
