Lusmse/syn_realDataset

Capstone VLA Datasets

LeRobot v2.1 datasets for a liquid pouring task on a KuavoV4Pro humanoid robot (34-DOF, bimanual). Used to benchmark GR00T N1.6, pi0.5, and Diffusion Policy under varying real-to-synthetic data ratios.

Mixture Configurations

Dataset	Type	Episodes
sim_mixed_40_60	unknown	500

Dataset Structure (LeRobot v2.1)

<dataset>/
├── meta/
│   ├── info.json
│   ├── tasks.jsonl
│   ├── episodes.jsonl
│   └── stats.safetensors
├── data/
│   └── chunk-*/
│       └── episode_*.parquet
└── videos/
    └── chunk-*/
        └── observation.images.*_episode_*.mp4

Pipeline

1. Teleoperation — ROS + Isaac Lab teleop (real demos)
2. Mimic annotation — Isaac Lab Mimic with Pink IK controller
3. Synthetic generation — Isaac Lab Mimic (~40% success rate)
4. Format conversion — HDF5 → LeRobot v2.1 (parquet + video)
5. Training — GR00T N1.6 · pi0.5 · Diffusion Policy

Benchmarking

Total dataset size held constant at ~600 episodes across all mixture configurations for controlled comparison.

---

Auto-generated by upload_datasets.py