Stanford AI and robotics professor Chelsea Finn talks with host Peter on The Robot Brains Podcast about helping AI systems adapt to changing conditions, with examples from language models and robots.
Finn says systems often fail when deployment data differs from training data, a problem called distribution shift. A robot trained to pick up a cup in one lab may not handle a different cup or setting, just as spam filters can fall behind changing messages. She describes approaches that improve robustness, enable quick adaptation, or let practitioners edit model behavior.
Finn describes model editors trained on examples of desired changes and examples that should remain unaffected. One method makes targeted rank-one weight updates; another stores edits in a memory and routes matching inputs to a separate model. Robots can also retry with new strategies, learn from their own data, and explore toward states they have encountered before.
Finn calls it single-life reinforcement learning when a robot uses prior offline experience but must keep learning in a new situation during one episode, without resets or human help. She says real environments rarely provide clear reward signals, and suggests sparse success-or-failure feedback or natural communication such as language and gestures. Video labeled with task instructions could help derive reward signals, though learned rewards can be exploited if they miss possible behaviors.
Finn says robot learning often needs human interventions every 10 seconds or, in a good case, every minute, and argues for more autonomous training. She also wants researchers to reuse and share data instead of starting each project from scratch. Her examples include the Bridge demonstration dataset, collected across about 10 environments, and a planned effort to gather data in closer to 1,000 environments.
Finn recalls that MAML began with the aim of adapting pretrained weights instead of training robots from scratch; an initial sinusoidal regression experiment worked on the first try. Looking ahead, she expects data quality and scale to matter, especially in robotics, where datasets remain small. Her advice is to build fundamentals in programming and debugging, then learn through projects and work that holds your interest.
Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.
Chelsea Finn of Physical Intelligence discusses building reliable, general-purpose robots, drawing on office coffee service, factory box assembly, and laundry-folding examples.
Chelsea Finn, cofounder of Physical Intelligence, presents at a Y Combinator event on building reliable, general-purpose robots for real-world work.
Chelsea Finn, co-founder of Physical Intelligence, discusses building general-purpose robots at a Y Combinator event.
Chelsea Finn, co-founder of Physical Intelligence and Stanford associate professor, discusses building general-purpose robots with the No Priors hosts.
Chelsea Finn, a Stanford professor, discusses how robots can generalize through broader training data and adapt to unfamiliar settings at a CMU Robotics Institute seminar.
Stanford University and Google Brain Professor Chelsea Finn presents her RI Seminar at Carnegie Mellon Robotics Institute on scaling data and algorithms for robot learning, introduced by host Wen.
