ethanCSL's picture
sfagen R04: exactly 100 synthetic episodes, evenly spread over the table
18f799b verified
|
Raw History Blame Contribute Delete
3.35 kB
metadata
license: other
license_name: robotwin-upstream
license_link: https://github.com/RoboTwin-Platform/RoboTwin/blob/main/LICENSE
task_categories:
  - robotics
tags:
  - LeRobot
  - robotwin
  - aloha-agilex
  - sim-free-augmentation
  - sfagen
configs:
  - config_name: default
    data_files: data/*/*.parquet

RoboTwin 2.0 stack_blocks_two (sfagen synthetic only)

LeRobot v3.0 dataset, 100 episodes / 31264 frames at 30 fps, robot aloha-agilex (14-D joint state/action, head camera cam_high + 2 wrist cameras, 640x480).

Source demos. The first 10 demo_clean episodes of stack_blocks_two in lerobot/robotwin_unified at revision 1287871839fa (global episode indices 24750-24759), split with LeRobot's split_dataset (videos re-encoded once, AV1 CRF 30).

Synthetic episodes. Made by sfagen (simulation-free augmentation; no physics simulator, no simulator state used) from the 10 source demos and the robot URDF only: the object (or the hand-over) is moved, the arm trajectory is retargeted by IK (pipeline v1: Cartesian blend, joint-space fallback near singular poses), and the head-camera image is re-rendered from the source frames. Every synthetic episode has exactly its source demo's length (1.000x).

  • families: object_workspace 100
  • shift from its source demo (object position, or hand-over location for interaction episodes): median 8.4 cm, max 11.6 cm
  • synthetic episodes per source demo: {'0': 14, '1': 7, '2': 14, '3': 8, '4': 8, '5': 6, '6': 7, '7': 17, '8': 6, '9': 13}
  • target positions: 3 cm lattice within 12 cm of the source object positions, kept when the object is visibly in the head camera (>= 50 % of its points, or as visible as in its source demo)
  • even spread (R04): exactly 100 synthetic episodes, chosen round by round over 3 cm cells (the emptiest cell first, ties by distance to everything already chosen; real demos count as occupied) so no area of the table is over-represented. Training-density evenness (episodes within 4 cm over the target lattice): CV 0.52 -> 0.40 vs the first release (r03, 195 synthetic episodes; still reachable at tag r03).
  • all candidates passed sfagen's kinematic/visual gates (speed/acceleration vs the source demos, IK, joint limits, collision, table, object placement); sfagen verify passed.
  • contains NO source episodes.
  • per-episode provenance (source demo, family, shift): meta/sfa_provenance_merged.json.

Image-quality warning (stack_blocks_two). good: the block is moved and the arm stays coherent; small arm fragments at the image edge. These are research data from a pipeline under development (the camera calibration is still 16-35 px off); inspect the videos before training on them.

Load (lerobot >= 0.4): LeRobotDataset("ethanCSL/robotwin_stack_blocks_two_lerobot_sim_aug_10", revision="v3.0"). Task string: "Set red block at the center, then stack green block over red block after placing it alongside." (see meta/tasks.parquet).

Provenance. sfagen code commit c7d5c01db1 (branch research/sfagen-simbench; local, not yet public). Upstream data: RoboTwin 2.0 (MIT-licensed code, https://github.com/RoboTwin-Platform/RoboTwin); the upstream LeRobot dataset declares no license, so the RoboTwin terms apply. Please cite RoboTwin 2.0 when using these data.