Reproduce “Discrete Policy: Learning Disentangled Action Space for Multi-Task Robotic Manipulation” on OpenResearch

About this paper

Kun Wu, Yichen Zhu, Jinming Li, Junjie Wen, Ning Liu, Zhiyuan Xu, Jian Tang

Researchers from Syracuse University, Midea Group, and Beijing Innovation Center of Humanoid Robotics developed Discrete Policy, a method that learns a disentangled action space for multi-task robotic manipulation. This approach, which combines a Vector Quantized Variational Autoencoder with a conditional latent diffusion model, achieves an 84% success rate on 5-task benchmarks and 66.3% on 12-task benchmarks for single-arm manipulation, along with over 65% success for bimanual tasks, outperforming prior diffusion-based policies.

Key insights

• Learning action patterns within a discrete latent space effectively disentangles multimodal action distributions into distinct, learnable 'skills' or 'action modalities'.
• A two-stage approach combining VQ-VAE for action discretization and a conditional latent diffusion model for generating discrete embeddings based on context improves multi-task policy learning.
• Disentangled action representations enable better generalization and facilitate the composition of learned skills for novel tasks.

Summary by alphaXiv.