shashuo0104/residual_copilot_models

Residual Copilot — Models

Policy checkpoints for Residual Shared Autonomy. Includes BC pilot models (DiffusionPolicy), RL residual copilot policies, and standalone autonomous baselines.

Contents

shared_autonomy_policies/
├── bc_teleop/                            # DiffusionPolicy pilots trained on teleop data
│   ├── GearMesh_bc_teleop/
│   ├── PegInsert_bc_teleop/
│   └── NutThread_bc_teleop/
├── bc_expert/                            # DiffusionPolicy pilots trained on expert data
│   ├── GearMesh_bc_expert/
│   ├── PegInsert_bc_expert/
│   └── NutThread_bc_expert/
└── residual_copilot/                     # RL residual copilot policies (RL-Games PPO)
    ├── {Task}_noisy_knn/nn/FactoryXarm.pth    # Trained with noisy kNN pilot
    └── {Task}_bc_teleop/nn/FactoryXarm.pth    # Trained with BC teleop pilot
autonomous_policies/                      # Standalone DiffusionPolicy baselines (no residual)
├── dp_teleop_data_match_succ/
│   └── gear_mesh_ckpt_30k/
├── dp_teleop_data_match_trials/
│   └── gear_mesh_ckpt_30k/
├── dp_residual_data_match_succ/
│   └── gear_mesh_ckpt_30k/
└── dp_residual_data_match_trials/
    └── gear_mesh_ckpt_30k/

Usage

Models are auto-downloaded at runtime via resolve_hf(). See the main repo for evaluation instructions.

# Evaluate residual copilot with noisy kNN pilot
python scripts/play.py --task GearMesh --pilot kNNPilot --copilot ResidualCopilot

Related

Repo	Type	Contents
residual_copilot_assets	dataset	Robot URDF/USD and object meshes
residual_copilot_data	dataset	Teleoperation training data (.npy)