This repository provides a minimal example of training a robotic manipulator in the Robosuite simulator. The agent uses a TD3 algorithm implemented in PyTorch.
Create a Python environment and install the dependencies:
pip install -r learning/requirements.txt
To start training run:
python learning/main.py
Training logs are written to the logs/
directory and model checkpoints are stored in tmp/td3/
.
After training, visualize the learned policy by running:
python learning/test.py
If you are on macOS, use the MuJoCo provided mjpython
interpreter when running any scripts or tests. For example:
mjpython learning/test.py # To be used only with has_renderer=True & has_offscreen_renderer=False
For test script with video recording, use python
instead of mjpython
.
This will load the latest checkpoints and render the environment while the agent interacts with it.
You can monitor training metrics using TensorBoard. Run the following command and open the displayed URL in your browser:
tensorboard --logdir="logs" --port 6006