This repository is the official submission of EPIC Lab for no external annotation track of SAPIEN ManiSkill Challenge 2021.
Please see environment.yml, we build our method on top of ManiSkill-Learn.
Please download ManiSkill demonstration dataset from here and store it in the folder training/data.
The training code is provided in training.
OpenCabinetDoor: run the shell command training/scripts/train_rl_agent/run_GAIL_door.sh
OpenCabinetDrawer: run the shell command training/scripts/train_rl_agent/run_SAC_drawer.sh
PushChair: run the shell command training/scripts/train_rl_agent/run_GAIL_chair.sh
MoveBucket: run the shell command training/scripts/train_rl_agent/run_SAC_bucket.sh
The evaluation code and the submisstion checkpoints of four tasks are provided in evaluation. You can use evaluate_policy.py from ManiSkill to run the model:
PYTHONPATH=YOUR_SOLUTION_DIRECTORY:$PYTHONPATH python mani_skill/tools/evaluate_policy.py --env ENV_NAME
For example, on OpenCabinetDoor, to evaluate the model:
PYTHONPATH=evaluation/Door:$PYTHONPATH python evaluate_policy.py --env OpenCabinetDoor-v0
Our trained models can be found at:
OpenCabinetDoor: Checkpoint
OpenCabinetDrawer: Checkpoint
PushChair: Checkpoint
MoveBucket: Checkpoint