RACER: Risk-sensitive Actor Critic with Epistemic Robustness

Kyle Stachowicz and Sergey Levine This is the release codebase for [RACER: Risk-sensitive Actor Critic with Epistemic Robustness](arxiv link).

Setup

Install the base requirements with:

conda create -n racer python=3.11
pip install -e jaxrl5[train] dmcgym .
pip install -r requirements.txt

Simulation

Run simulation with the following command:

python scripts/sim/train_online_states.py --config scripts/sim/configs/distributional_limits_config.py --world_name <flat|bumpy> --cvar_risk 0.9

Real robots

Getting things set up from scratch on a new hardware platform is somewhat involved. We use the robot hardware and training workstation setup from FastRLAP, which requires two devices: local compute on the robot (Jetson Orin NX) and a workstation or desktop with GPU for training. See the offroad-robot-ros directory's instructions for more information.