VLA-RL-Tasks

This repository is for submitting custom-designed simulation tasks for RL training VLA models. The automated CI will verify the correct implementation under randomly sampled actions, but it will not validate the reward code.

Currently only supporting Maniskill CI!

Instructions

1. Create a folder under src/tasks named after yourself, with your name properly spelled and capitalized (e.g., JunhaoChen).
2. Define your tasks in .py files within your folder.
3. Check the correctness of your custom tasks by running src/local_runner.py, ensuring imports and gym.make parameters are adjusted accordingly. Do not modify this file in the repository before submitting pull requests. Instead, use your own local version for testing.
4. After submitting a pull request, the repository will trigger a CI process that automatically generates trajectory videos with randomly sampled actions. Once the CI passes, pull the repository again to sync the generated videos and review them for any potential bugs.
5. Lastly, update the table below to track your progress.

Task List

Task ID	Task Name	Env Name	Task Instruction	Sim Env	Name
0	Card Stacking (Example)	CardStack-v1	Stack the cards onto each other.	Maniskill	Junhao Chen
1	...	...	...	...	...