Reproduce “RoboSSM: Scalable In-context Imitation Learning via State-Space Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Youngju Yoo, Jiaheng Hu, Yifeng Zhu, Bo Liu, Qiang Liu, Roberto Martín-Martín, Peter Stone
Researchers at The University of Texas at Austin developed RoboSSM, an in-context imitation learning framework that utilizes State-Space Models (SSMs) as its backbone to overcome the scalability and generalization limitations of Transformer-based methods in robotics. This framework processed demonstration prompts up to 16 times longer than those it was trained on, achieving superior performance on unseen tasks and long-horizon tasks, while maintaining linear-time inference.
Key insights
- State-Space Models provide an efficient and scalable alternative to Transformers for In-Context Imitation Learning, offering linear-time inference for long sequences.
- The inherent online regression formulation of Longhorn contributes to strong extrapolation capabilities, enabling robust generalization to out-of-distribution prompt lengths and task complexities.
- Strategic modulation of the SSM's weighting vector (β-scaling) can further optimize the model's ability to leverage in-context demonstrations for novel task inference.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.