RoboSSM: Scalable In-context Imitation Learning via State-Space Models

RoboSSM introduces a novel approach to in-context imitation learning (ICIL) utilizing state-space models (SSM) instead of Transformers, specifically employing Longhorn for its linear-time inference and enhanced extrapolation capabilities. This method demonstrates significant improvements in generalization for unseen and long-horizon tasks on the LIBERO benchmark, surpassing traditional Transformer-based ICIL methods by effectively managing longer prompts during testing. The findings highlight SSMs as a scalable and efficient alternative for practitioners looking to enhance ICIL performance in robotics applications.