a3c-mujoco

Simulated target reaching tasks using the MuJoCo physics engine. The setup is adapted from [1] end-to-end learning setup for solving pixel-driven control of Jaco arm where learning is accomplished using Asynchronous Advantage Actor-Critic (A3C)[2] method with sparse rewards.

Usage:

Run with python main.py <options>.

Dependencies:

Python 3.5.2
NumPy
mujoco-py 1.50.1
OpenAI Gym
PyTorch
Plotly

Note:

Obtain a 30-day free trial on the MuJoCo website or free license if you are a student.

Results

Acknowledgements

@kaixhin for ACER
@ikostrikov for pytorch-a3c

References

[1] Sim-to-Real Robot Learning from Pixels with Progressive Nets
[2] Asynchronous Methods for Deep Reinforcement Learning