- An implementation of Short Horizon Actor Critic (Xu; 2022) writen in Jax
- Simulation using the Mujoco MJX simulator
Run Time: 1 min jit, 2 min training
Known Issues: For some random seeds, you get drift in the cart position.
Run Time: 1 min jit, 2 min training
Known Issues: As seen in the rewards figure, training can be unstable.
- Having great difficulty applying SHAC to get Anymal to walk with default 32-bit precision. (See Mujoco for an example with 64-bit precision)
- Hypothesis: it's because quadruped gait is very contact-rich, leading to uninformative gradients.
32-step rollout. Ground flashes red when step jacobian is greater than 10e2.
- pip install -r requirements.txt
- Add the parent folder of this repository to your PYTHONPATH environment variable.