Introduction

This repository is the official implementation of Silver-Bullet-3D Solution for SAPIEN ManiSkill Challenge 2021

Requirement:

• PyTorch 1.8.0+
• Python3.7
• CUDA 10.1+

Other requirements please refer to environment.yml

Clone the repository:

git clone https://github.com/caiqi/Silver-Bullet-3D

No Interaction Track

Data preparation

• Download ManiSkill dataset from here
• Compress the data with compress_data.sh

Training

Training code is provided in No_Interaction/training folder. For example, to train MoveBucket model, using following script:

CONFIG_NAME=bucket/1225_bucket_ensemble_v1
SEED=1345
CUDA_VISIBLE_DEVICES=0 python -m tools.run_rl configs/${CONFIG_NAME}.py --gpu-ids=0 --seed ${SEED} --work-dir ${CONFIG_NAME}

For final submission, we ensemble multiple models with different network architecture and random seed. All configs are provided in configs folder.

Evaluation

Evaluation code and checkpoints are provided in No_Interaction/evaluation. For example, to evaluate the pre-trained models on MoveBucket, use ManiSkill official evaluation code:

PYTHONPATH=No_Interaction/evaluation/bucket_track1:$PYTHONPATH python evaluate_policy.py --env MoveBucket-v0 --level-range "0-300"

evaluate_policy.py is from ManiSkill repo. Checkpoints can be downloaded from release page.

Task	Models
OpenCabinetDoor	Checkpoint
OpenCabinetDrawer	Checkpoint
MoveBucket	Checkpoint
PushChair	Checkpoint

No Restriction Track

The training and evaluation code is the same.

Task	Code
OpenCabinetDoor	user_solution_door.py
OpenCabinetDrawer	user_solution_drawer.py
MoveBucket	user_solution_bucket.py
PushChair	user_solution_chair.py

Acknowledgements

Thanks the contribution of ManiSkill-Learn and awesome PyTorch team.