We provide a set of common baselines from the Robot Learning literature to help create and develop new solutions for BEHAVIOR. Our baselines include imitation learning and reinforcement learning, both in the original low-level action space of the benchmark, and making use of our provided set of action primitives based on sampling-based motion planning. Here, we briefly describe the baselines you can find in this repository.
Imitation Learning Baseline
There are two BC agents in
BCNet_rgbp: A BC agent that uses an RGB image (128x128) and proprioception feedback (20) as state space.
BCNet_taskObs: A BC agent that uses task observations (456) and proprioception feedback (20) as state space.
Details about state information can be found here. Feel free to include additional state information such as depth, instance segmentation, etc.
Both agents are based on the BehaviorRobot, you can find more details about its action space here. Please note that the action vector in human VR demo contains two additional dimensions that correspond to the hand reset action in VR: an action that teleports the simulated hands to the exact pose of the VR hand controller when they have diverged. These actions are not used by the agents but are necessary to understand the demos.