
Learning-enabled Adaptation to Evolving Conditions for Robotics
Somrita Banerjee of Stanford University presents this talk in the AA289 Stanford Robotics and Autonomous Systems Seminar, addressing how robots can adapt when real-world conditions diverge from training data. She covers two approaches. The first handles novel task instances that share underlying structure, using passive parameter adaptation that augments physics-based models with learned components while keeping physical parameters interpretable, so engineers can monitor how the adaptation unfolds. The second introduces a framework for active adaptation, where a model tracks its own performance, identifies a diverse set of uncertain inputs, and uses them for periodic fine-tuning across the system's full deployment lifecycle. The talk is grounded in Banerjee's research at Stanford's Autonomous Systems Lab and aimed at viewers interested in machine learning applied to autonomous robotics operating in unstructured or unpredictable environments.