Continual/lifelong learning for long-term robotic operation

Establish practical continual or lifelong learning methods for robotics that enable robots to operate over long periods while adapting to new tasks and environments without sacrificing previously acquired capabilities, thereby overcoming catastrophic forgetting during sequential task acquisition.

Background

The paper studies continual learning for vision-language-action (VLA) models in robotics, where robots must acquire new skills over time without losing previously learned abilities. Naive sequential fine-tuning of pre-trained VLAs causes catastrophic forgetting, making it unsuitable for long-term deployment.

Existing approaches often rely on experience replay, regularization, or architectural changes, but they have practical drawbacks such as storage requirements for exemplars, reliance on task identifiers, or limited scalability for long task sequences. This motivates the identification of continual and lifelong learning in robotics as an open challenge, which the paper addresses with the proposed CLARE framework.

References

This long-term adaptability, known as continual or lifelong learning, remains an open challenge in robotics despite decades of research.

— CLARE: Continual Learning for Vision-Language-Action Models via Autonomous Adapter Routing and Expansion  (2601.09512 - Römer et al., 14 Jan 2026) in Introduction, Section 1

While the online actor responds to current observations, handling changes in task goals or newly introduced instructions during an episode remains unexplored.

— Fast-VLA: You Only Think Once for Robot Manipulation  (2610.21389 - Gong et al., 15 Oct 2026) in Appendix, Section “Limitations”

The open problem is cumulative improvement with bounded verification costs and without sacrificing service continuity or safety.

— Physical AI: A Survey of Models, Learning, Systems, and Open Challenges  (2610.21171 - Gao et al., 15 Oct 2026) in Section 10.4.4, Sustained Operation

However, current approaches mainly focus on memory storage and retrieval, while autonomous organization, knowledge consolidation, and adaptive updating remain open challenges.

— From Sliding Windows to Self-evolving Lifelong Memory: A Dual-perspective Survey of Memory Paradigms for Long-Horizon Robotics VLA Models  (2610.23319 - Wang et al., 16 Oct 2026) in Section 4.2, Deployment-time Persistent Memory

However, fully autonomous self-evolving memory has not yet been demonstrated in current VLA systems [1].

— From Sliding Windows to Self-evolving Lifelong Memory: A Dual-perspective Survey of Memory Paradigms for Long-Horizon Robotics VLA Models  (2610.23319 - Wang et al., 16 Oct 2026) in Section 4.3, Full-lifecycle Multimodal Self-evolving Memory

Learning continually from a non-stationary stream of experience, as humans do, remains a major open challenge.

— Is this machine playing?  (2610.07130 - Cloos et al., 5 Oct 2026) in Section 5.4, “Continual and active learning in externalized artifacts”