RoboMME-Plus
Collection
3 items • Updated
50 episodes · 20 fps · 2 cameras · 256×256 h264
RoboMME-OOD training data (LeRobot v3)
50 successful training episodes · 16 tasks · 56,462 frames · 20 Hz. Three examples (easy, medium, hard) per task, except MoveCube with one example for each of five methods. Train seeds are disjoint from the 800 test episodes.
Source and seeds
Matching RoboMME-OOD code: d8a7e67852031d0f6e32e683873966c969601773.
Training seeds are in the RoboMME repo at
src/robomme_ood/env_metadata/train/record_dataset_<Task>_metadata.json.
Record k corresponds to H5 episode_<k>. This dataset also includes
train_seeds.csv, train_plan.json and metadata/train/.
Files and conventions
meta/, data/ and videos/ form a LeRobot v3 dataset with one Parquet file and two camera videos per episode. meta/source_episodes.json maps episode indices to tasks and seeds.
Observations precede actions. Video-conditioned episodes include demonstration
frames followed by execution; is_video_demo distinguishes them. Angles are radians
and positions metres. VideoPlace examples finish with an open-gripper upward
retreat of at least 5 cm. LeRobot H264 videos are lossy; H5 stores the original RGB.
See the field reference.
Validation
All 50 execution trajectories pass public-evaluator physics replay with zero
joint trajectory difference. All 117 format checks and both dataset-loader checks
passed. manifest.json lists file sizes and SHA-256 checksums;
validation.json summarizes the checks and source revision.
- Downloads last month
- 7