dm_control
dm_control copied to clipboard
How to define new task ?
We would like to simulate a custom robotic manipulation task environment and then use it in a reinforcement learning framework. It is possible to use dm_control components to do that and is there any guidelines on how to proceed ?
I used the tasks in manipulation, locomotion and suite to learn the API. They're extremely well-written and should cover a lot of uses cases.
Thanks @kevinzakka! I find there is another way to create custom task in dm_robotics https://github.com/deepmind/dm_robotics/blob/main/py/moma/README.md It uses dm_control as well. I’m wondering which is the best way to go ?
@jrabary The dm_robotics repo builds on top of dm_control and uses library-specific things like agentflow, base/subtask abstractions, etc. It has more features but also a much steeper learning curve. I'd recommend you start with dm_control as it should cover most RL use cases.