README.md

metadata

license: cc-by-4.0
task_categories:
  - robotics
tags:
  - LeRobot
  - dynamic-robot-bench
  - conveyor
  - scripted-expert

dynamic_robot_bench_dr_scripted_10k

10,000 scripted-expert demonstrations across all 100 dynamic task families of dynamic-robot-bench — a conveyor-belt dynamic-manipulation benchmark (Franka Panda + wrist camera, ManiSkill 3 / SAPIEN GPU sim). One LeRobot v2.1 dataset: 100 episodes per family, success-filtered, language-prompted per episode.

Collection configuration (identical for every family)

• Scripted expert with per-step auto-derived speed caps, recorded as joint-position actions in the droid_panda schema (3 RGB views: 2 exterior + wrist, 224×224, 20 fps).
• Full domain randomization per episode: belt speed uniform 0.10–0.50 m/s, stratified over 5 equal-width bins (20 episodes per bin per family), belt direction, task-specific spatial/physical knobs (object yaw, lateral offsets, spring stiffness, ...), expert velocity SDE, and scene-background DR (workspace texture, lighting, camera jitter, clutter) resampled every 20 episodes. Families with an object catalog resample object identity as well.
• Success-filtered: only episodes whose task-specific success predicate holds at termination are kept. Every family's expert passes >=0.90 success under this full randomization (verified per family at multiple seeds before collection).

Layout

• data/chunk-XXX/episode_NNNNNN.parquet — one file per episode, images inline.
• meta/episodes.jsonl / meta/episodes_stats.jsonl / meta/tasks.jsonl / meta/info.json — standard LeRobot v2.1 metadata.
• ranges.json — maps each of the 100 gym task ids to its contiguous episode range (episodes are ordered by family, alphabetical by gym id).

Asset attribution

Rendered scenes may include third-party 3D assets:

• Google Scanned Objects (CC-BY 4.0) — Google LLC. https://app.gazebosim.org/GoogleResearch/fuel/collections/Scanned%20Objects%20by%20Google%20Research
• YCB Object and Model Set (CC-BY 4.0) — Calli et al.
• Procedural primitive assets generated by the benchmark itself.

This dataset is released under CC-BY 4.0.