Skip to content

Latest commit

 

History

History
49 lines (30 loc) · 1.44 KB

README.md

File metadata and controls

49 lines (30 loc) · 1.44 KB

VQ-ACE: Efficient Policy Search for Dexterous Robotic Manipulation via Action Chunking Embedding

The code framework is adapted from srl_il

The VQACE-MPC implementation is in vqace_mpc

The VQACE-RL implementation is in the vqace_rl branch of faive_gym_oss

Get started

The following is tested on Ubuntu22.04

conda create -n vq_ace python=3.11 # or other versions of python
conda activate vq_ace

git clone [email protected]:srl-ethz/vq_ace.git
cd vq_ace
pip install -e .

Training

Train the action chunk embeddings without vector quantization.

python3 scripts/run_pipeline.py --config-name=train_embed_act

Train the action chunk embeddings with vector quantization.

python3 scripts/run_pipeline.py --config-name=train_embed_vq_act

Train the action chunk embeddings with vector quantization, but without conditions

python3 scripts/run_pipeline.py --config-name=train_embed_vq_act

🙏Acknowledgement

  • The robomimic tasks and observation encoders are adapted from Robomimic
  • The linear normalizer implementation is adapted from diffusion policy
  • The vector quantize implementation is adapted from vq_bet_officia