Reproduce “DSSP: Diffusion State Space Policy with Full-History Encoding” on OpenResearch

About this paper

Zhiyuan Guan, Jianshu Hu, Han Fang, Yunpeng Jiang, Yize Huang, Shujia Li, Xiao Li, Yutong Ban

A Diffusion State Space Policy (DSSP) with full-history encoding leverages State Space Models (SSMs) to efficiently compress long observation histories, addressing temporal ambiguities in robot manipulation tasks. The approach achieved a 12.8% relative improvement in success rate on the RoboTwin 2.0 benchmark and a 133.3% relative improvement in real-world tabletop tasks compared to a leading baseline, while using a smaller model and reducing inference latency by 45.4%.

Key insights

• State Space Models (SSMs) like Mamba provide an efficient and effective backbone for encoding long observation histories in robot manipulation, offering linear-time complexity and compact memory usage superior to Transformers for this task.
• Explicitly conditioning policies on a full, learned history context significantly reduces imitation loss and improves performance in long-horizon, memory-dependent tasks by resolving temporal ambiguities.
• An auxiliary dynamics-aware loss function can effectively guide the learning of a history context that is not only compact but also predictive of future states, ensuring it retains action-relevant information.
• A hierarchical conditioning strategy that combines high-level, long-term history with recent local observations, coupled with timestep-decoupled noise injection, stabilizes and enhances the diffusion-based action generation process.

Summary by alphaXiv.