Chelsea Finn, co-founder of Physical Intelligence, discusses building general-purpose robots at a Y Combinator event.
Finn says solving each robotics application separately requires custom hardware, software, and movement strategies, making it difficult to bring robots into everyday settings. Physical Intelligence is developing a model intended to let different robots handle many tasks. She argues that large datasets are necessary, but industrial, video, and simulated data each miss important kinds of real-world variety or direct experience.
Finn describes training a robot to unload a dryer and fold clothes, a long task with variable garments and many chances for small errors to derail progress. After months of poor results, the team improved performance by pretraining on broad robot data and then fine-tuning on consistent, curated demonstrations. A roughly 3-billion-parameter vision-language model further improved speed and consistency; the same training approach also transferred to tasks such as cleaning a table, scooping coffee, and making coffee on another robot.
To test robots in new places, the team collected tidying data across more than 100 rooms and combined it with other robot, web, and instructional data. In three unfamiliar rented homes, robots followed requests to put away dishes, clean spills, and tidy bedrooms. Finn reports that broader environmental coverage improved performance, while success remained around 80%, with failures involving drawers, thin objects, and getting stuck.
Finn describes a hierarchical system that turns broad requests into smaller spoken instructions for a robot. The team used language models to create hypothetical human prompts from existing robot demonstrations, reducing the need to collect every interaction directly. Robots demonstrated requests such as making a vegan sandwich without pickles and adapting when a user changed or corrected an instruction.
Finn says real-world robot learning still needs more research, reliable systems, and contributions to shared data and models. In the Q&A, she highlights fast on-robot inference and large-scale training infrastructure, and says online robot experience could improve post-training speed and success. She sees real robot data as essential, with simulation useful for evaluation and robot attempts useful for learning.
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 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.
