Chelsea Finn, cofounder of Physical Intelligence, presents at a Y Combinator event on building reliable, general-purpose robots for real-world work.
Finn describes Physical Intelligence’s progress on tasks such as folding laundry, washing pans, and making food, then focuses on what makes robots useful beyond impressive demos. Unlike recommendation systems, robots act directly on the physical world, so they need to make fewer mistakes and work autonomously for long stretches. She uses making espresso as an example: it requires precise force, careful handling of liquid, and reliable timing.
Finn argues that reaching high reliability requires repeated improvement, but manual data collection and tuning can become burdensome. Physical Intelligence’s approach uses human interventions to steer robots away from unproductive failures, plus a general-purpose value function trained on robot videos to estimate progress across tasks. In examples including latte making and box building, reinforcement learning improved throughput by about 2×; the espresso policy exceeded 90% success and ran for 13 hours during evaluation.
Finn says most robotics foundation models act on current observations without memory, which limits their ability to track progress through multi-step work. Physical Intelligence combines efficient short-term video memory with longer text summaries of events. With memory spanning roughly 10 to 15 minutes, a robot completed a varied kitchen-cleaning task autonomously, including wiping counters, putting mustard away, and washing dishes.
Finn places robotics at an early stage of the shift from task-specific systems to general-purpose models. Physical Intelligence’s π₀.₇ model was trained on diverse robot demonstrations, rollouts, human videos, and web data, with prompts that include memory, subtasks, and metadata about data quality. The model matched or exceeded fine-tuned specialists on several tasks and showed transfer to a rare air fryer and to folding clothes on a different robot platform without task-specific training data.
Finn says robotics is entering an era of capable general-purpose models, while noting that deployment will spread more slowly than ChatGPT because robots must be physically installed. She expects real-world robot experience to be essential training data, alongside useful sources such as human videos and web content. In the Q&A, she recommends that small teams start with generalist policies and fine-tune them, and says software engineers can enter robotics by joining a company or experimenting with accessible hardware and open-source models.
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, 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.
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.
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.
