aaroncaozj/BlockPAP-v0_Mix

--- configs: - config_name: default data_files: - split: train path: - data/chunk-000/*.parquet - data/chunk-001/*.parquet language: - en license: apache-2.0 pretty_name: BlockPAP-v0 Mix --- > BlockPAP-v0_Mix > `LeRobot v2.1` mixed dataset with 2 camera views for `BlockPAP` (real + MimicGen trajectories). <br> # ❗️ State Alignment between Simulation and Real Robot On the **Franka Research 3**, the end-effector states in the real robot dataset contain systematic offsets compared to the simulation due to calibration differences. | State Index | Description | Offset (Real vs. Sim) | | ----------- | ---------------------------- | --------------------- | | `state[2]` | End-effector z position | **+10 cm** | | `state[5]` | End-effector yaw (z_euler) | **−45°** | ## Training Requirement To ensure consistent data distributions between simulation and real demonstrations: * When loading **simulation data**, add the corresponding constant bias to `state[2]` and `state[5]`. * This aligns the state distributions across both sources and enables smoother deployment on the real robot. ## Simulation Rollout * If performing **policy rollout in simulation**, apply the same bias to the input states before feeding them to the model. This keeps the evaluation distribution consistent with the training distribution.