Algorithms for Data-efficient Continual Robot Learning

<p dir="ltr">The deployment of intelligent robotic agents in complex, non-stationary human environments—such as collaborative manufacturing, autonomous driving, and long-term navigation—hinges on their ability to learn continually from a dynamic stream of real-world data. Traditional machine learning paradigms, which rely on offline training with static datasets, are fundamentally ill-suited for these settings. They are plagued by critical limitations, most notably catastrophic forgetting, where new knowledge overwrites previously learned skills; profound data inefficiency, a particularly acute problem in robotics where data collection is expensive and time-consuming; and the challenge of modeling and controlling complex nonlinear dynamics with adaptability and generalizability from limited data. </p><p dir="ltr">This thesis presents a systematic investigation into a new class of learning algorithms designed to overcome these barriers, enabling robust and data-efficient continual learning for robots. The contributions are organized into three syner?gistic pillars. The first pillar, Data-Efficient Optimization, introduces novel algorithms for both online and offline learning. We develop a family of Extended Kalman Filter (EKF)-based optimizers that leverage second-order information to achieve superior convergence for online adaptation. We also improve adaptive gradient methods for offline pretraining by introducing a data-driven approach to initialize optimizer states, enhancing stability and performance. The second pillar, Continual Learning with Memory Mechanisms, addresses catastrophic forgetting through bio-inspired memory architectures. This includes a feedforward compensation strategy that proactively uses critical past experiences to improve adaptation and the BioSLAM dual-memory system, which explicitly manages short-term plasticity and long-term knowledge consolidation for lifelong place recognition. The third and culminating pillar, Models for Efficient Robot Learning, presents novel frameworks for learning generalizable system models for control. We pioneer the continual learning and lifting of Koopman dynamics, a breakthrough approach for linearizing high?dimensional, nonlinear systems like legged robots from streaming data. This pillar also investigates methods for estimating neural network robustness via Lipschitz constants, a crucial step for deploying learned models in safety-critical applications. </p><p dir="ltr">Collectively, this body of work provides a foundational framework for creating truly adaptive, resilient, and intelligent robots. By developing solutions for data-efficient optimization, memory-augmented learning, and continual dynamics modeling, this thesis paves the way for the safe and effective deployment of autonomous systems in the open world.</p>