Reproduce “Social-Mamba: Socially-Aware Trajectory Forecasting with State-Space Models” on OpenResearch

About this paper

Po-Chien Luan, Wuyang Li, Yang Gao, Alexandre Alahi

Social-Mamba introduces a framework for human trajectory forecasting that leverages State-Space Models (SSMs) to efficiently model complex social interactions in crowded environments. The system achieves state-of-the-art accuracy on datasets like NBA-Full while significantly reducing computational parameters by 75% and GFLOPs by 54% compared to previous Transformer-based approaches.

Key insights

• The Cycle Mamba block effectively integrates bidirectional context at the state level within SSMs, allowing for sophisticated dependency modeling with fewer parameters compared to standard bidirectional approaches.
• Decomposing complex social interactions into specialized, ego-centric sequential scans and dynamically fusing their outputs enables SSMs to adapt to unstructured social dynamics while preserving spatial context and varying influences.
• State-Space Models, when appropriately adapted and structured, can provide a computationally efficient and scalable foundation for multi-agent social reasoning, outperforming attention-based Transformers in efficiency without sacrificing accuracy.

Summary by alphaXiv.