pretty_name: ctr-sortblocks-100ep-sequential
tags:
- lerobot
- robotics
- simulation
language:
- en
ctr-sortblocks-100ep-sequential
100 native RoboTwin Blocks Ranking RGB-CTR expert episodes in LeRobot v3 format. This is synthetic expert data, not a learned-policy benchmark.
Task
The left arm places red then yellow; the right arm places blue then green. Initial red/blue and yellow/green poses are mirrored across the arm bisector. Final world-x order is red, yellow, blue, green. Group centers are 0.52 m apart. Every physics step checks a conservative minimum 0.10 m cross-arm/actor clearance. The lowest recorded clearance in this dataset is 0.185744 m.
Sampling
Episodes 0–49: physical left arm first. Episodes 50–99: physical right arm first. The delayed arm starts exactly when the first arm finishes.
All four companion datasets use the same 100 accepted initial scenes and frozen per-scene joint/gripper controls. Timing changes do not change trajectory speed. Mixed reuses the selected source episodes; it introduces no new simulation scenes. Source scene IDs, control hashes and exact offsets are in meta/retime_manifest.json.
Contents
- Episodes: 100; training rows: 91191; FPS: 25; physics: 250 Hz.
- Robot: Aloha AgileX; observation.state: 14 measured joint/gripper values.
- action: 14 commanded joint/gripper targets at the next recorded observation.
- Images:top, left_wrist and right_wrist, 320×240 RGB H.264 video.
- Wrist preset: centered_fovy90, including calibrated optical-axis mounts.
- Fixed wider head intrinsics: fx=fy=260 px, cx=160, cy=120.
- Last observations are verification/next-target material, not extra action rows.
IdleMask
This non-CTR dataset has no IdleMask columns. Training with this dataset does not enable IdleMask.
Statistics and provenance
Numeric publication statistics use all valid published numeric rows, with exact global linear quantiles computed after promotion to float64. This includes state, action, timestamps and timing metadata. Image quantiles are omitted; native LeRobot sampled image mean/std estimates are retained. There is no episode-quantile averaging, padding, or mask-based exclusion. Training-specific delta/action-window normalization is a separate transform.
meta/source_programs contains the frozen controls and initial scenes. .ctr-revision.json identifies the generated scene/control manifest; the immutable Hugging Face commit is recorded in the external publication receipt. SHA256SUMS pins the uploaded files. All episodes passed task success, camera, control-integrity and timing checks before publication.