scripts

Setting up

1. Follow the installation instructions for AgentHive.
2. Download FK-v1(expert) and DAPG(expert) dataset from the RoboHive dataset collection - RoboSet.

Behavior Cloning Running Instructions

sim_backend=MUJOCO MUJOCO_GL=egl python bc/run_bc_h5.py \
                                    encoder = <visual-encoder> \
                                    cam_name = <camera-name> \
                                    env_name = <env-name> \
                                    from_pixels = True \
                                    data_file = <path-to-dataset>

Currently, three visual encoders are supported: VC1, R3M, RRL. To use the largest model variant of each one set encoder=vc1l/r3m50/rrl50.
To run the experiments using privileged state or proprioceptive-only information, set from_pixels=False and encoder=state/proprio, respectively.

Results

For each of the visual baselines, the results are averaged over 3 camera angles (except for the Robel Suite, where all the camera angles are used to avoid partial observability), 3 seeds, and 25 evaluation trajectories.

Benchmark Suite	Dataset Type	Camera Angles	Seeds
Kitchen (FK-v1)	FK-v1(expert)	left_cam,right_cam,top_cam	1,2,3
Kitchen (FK-v1)	FK-v1(human)	left_cam,right_cam,top_cam	1,2,3
Hand Manipulation Suite (HMS)	HMS(Human)	view_1,view_4,vil_camera	1,2,3
Robel Suite	ROBEL(Expert)	[A:headCam,A:leftCam,A:tailCam,A:rightCam]	1,2,3