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

Researchers introduce RoboSSM, a new in-context imitation learning framework that replaces Transformers with state-space models (SSMs) for robotic task learning. The approach demonstrates superior performance on long-context prompts and achieves better generalization to unseen tasks compared to Transformer-based methods, establishing SSMs as a viable alternative backbone for robot learning systems.

RoboSSM addresses a critical limitation in current robotic imitation learning systems: Transformers struggle with computational efficiency and fail to extrapolate beyond training sequence lengths. This research validates state-space models as a superior alternative for in-context imitation learning, leveraging Longhorn's linear-time inference and strong extrapolation properties. The work is significant because it challenges the Transformer-dominant paradigm in machine learning, demonstrating that specialized architectures can outperform general-purpose models in specific domains.

The broader context reveals growing recognition that Transformers' quadratic complexity and limited generalization create bottlenecks for deployment-time adaptation in robotics. Prior ICIL methods faced practical constraints when handling demonstration sequences beyond their training distribution, limiting real-world applicability. RoboSSM's success on the LIBERO benchmark—a comprehensive robotics evaluation suite—indicates that SSMs' mathematical properties align well with robotic decision-making requirements.

For the AI and robotics industry, this research accelerates the shift toward more efficient, specialized architectures. Organizations developing robotic systems benefit from lower computational requirements at inference time and improved task generalization. This opens opportunities for deploying sophisticated robotic learning on resource-constrained devices and enables more flexible few-shot adaptation in production environments.

Looking ahead, the validation of SSMs in robotics may inspire broader architectural exploration across AI domains. The reproducibility of results (code availability) strengthens the foundation for subsequent research building on this approach. Future work likely explores hybrid architectures combining SSM strengths with other advances, while scaling these methods to increasingly complex robotic tasks and multi-agent scenarios.

• →State-space models outperform Transformers for in-context imitation learning with linear-time inference and superior long-context handling.
• →RoboSSM achieves better generalization to unseen and long-horizon robotic tasks than existing Transformer-based ICIL methods.
• →SSMs demonstrate strong extrapolation capabilities, enabling effective learning from longer demonstration sequences than those seen during training.
• →Linear-time inference of SSM-based approaches reduces computational overhead for real-time robotic deployment and adaptation.
• →Results validate SSMs as efficient, scalable backbones for robotic learning, challenging Transformer dominance in the field.