Chelsea Finn

Chelsea Finn studies how robots and other agents can develop broadly intelligent behavior through learning and interaction. She is an assistant professor in Computer Science and Electrical Engineering at Stanford University and cofounder of Physical Intelligence, a role she began in 2024. Her Stanford lab, IRIS, studies intelligence through robotic interaction at scale; her research includes collaborative work on meta-learning and generalist robot policies.[2][3][4][6]

At Startup School 2026, Finn argued that useful robotics requires sustained autonomous operation. A recommendation system can leave the final decision to a person; a robot directly changes the physical world and therefore needs to make fewer mistakes. Her example was espresso preparation: inserting a portafilter requires precise, forceful control, while moving a full cup requires smooth motion and careful timing. Performing the task once is only the beginning; usefulness depends on repeating it reliably without frequent supervision.[7]

Finn described a reinforcement-learning recipe designed to use scarce robot experience efficiently. Humans intervene to stop dead-end attempts and demonstrate recovery—for example, separating two cardboard boxes that a robot mistakenly picked up together. A general-purpose value function learns from robot videos to estimate progress toward success across tasks, reducing the need for repeated attempts at each individual task. The resulting experience is used to improve the policy.[7]

In the talk published August 12, 2026, Finn reported approximately twice the throughput from the reinforcement-learning stage in the experiments presented, over 90% success on espresso preparation, and a 13-hour run of the latte-making policy. In the latte workflow, a person steamed the milk. She kept the limitations explicit: “Even with this improvement, the robot still makes mistakes. It’s also still slower than people.”[7]

For tasks that require tracking a sequence of completed steps, Finn described memory at two timescales: roughly ten seconds of efficiently computed video history, and compressed textual summaries covering longer periods. She reported a 10–15-minute autonomous kitchen-cleaning workflow that included wiping and drying a counter, putting objects away, and washing dishes. This differs from repeatedly making espresso: the robot must remember which steps it has already completed.[7]

Finn identified two goals for general-purpose robotics: a model that works out of the box, and compositional generalization, in which familiar skills, objects, and robot platforms can be combined in new ways. She reported that a single π0.7 model matched or exceeded task-specific specialists in the comparisons she presented. Its training combined demonstrations, robot attempts, human video, and web data. Prompts supplied context such as memory, immediate subtasks, and metadata about data quality; she reported that this metadata helped the model benefit even from lower-quality training examples.[7]

Finn’s own hands-on example was an initial pinwheel-assembly policy she said she personally trained while working on π0.7. The demonstrations consistently placed the pin in the right gripper and the paper in the left. After an error reversed their positions, she recalled that the robot inserted the pin using its left gripper instead. Finn said this reversed-hand strategy appeared in neither pretraining nor post-training data, and interpreted the recovery as a transfer of behavior between hands.[7]

Her air-fryer example carried a qualification. The team initially thought its dataset contained no air-fryer demonstrations, but later found three episodes. Finn presented the robot’s ability to open the appliance, place a sweet potato inside, and close it as generalization from a scarcely represented appliance. She expected the three episodes were unimportant, but did not report an experiment removing them.[7]

Finn argues that human video and web data help robot learning, while experience on the robot’s own platform remains critical: “I think that there’s no substitute for the robot experience itself.” She expects future datasets to contain substantial autonomous experience from robots attempting real-world tasks. She also expects distribution to be slower than ChatGPT’s because deploying a physical model requires deploying physical hardware.[7]

An earlier step in this work was π₀, announced October 31, 2024, which Finn coauthored with 23 collaborators. The team augmented a pretrained vision-language model with continuous action outputs through flow matching, training it on diverse robot data and then specializing it with high-quality task data. The approach combined semantic understanding from Internet-scale pretraining with embodied experience for dexterous control. The team demonstrated post-trained policies for tasks including laundry folding, table bussing, and cardboard-box assembly.[6]

Finn’s work on adapting to new tasks predates Physical Intelligence. In 2017, she, Pieter Abbeel, and Sergey Levine introduced Model-Agnostic Meta-Learning, or MAML. The method trains model parameters across tasks so that a small amount of new-task data and a few gradient steps can produce good generalization. As the authors put it, “In effect, our method trains the model to be easy to fine-tune.” It applies to models trained with gradient descent, including classification, regression, and reinforcement learning.[4]

That same year, Finn, Tianhe Yu, Tianhao Zhang, Pieter Abbeel, and Sergey Levine published One-Shot Visual Imitation Learning via Meta-Learning. Their method used prior-task training to enable adaptation from a single visual demonstration, with raw pixels as inputs. The authors reported learning new tasks on simulated and real robots. The single demonstration was the input for learning a new task after prior training, rather than the robot’s entire learning experience.[5]

Finn completed a BS in electrical engineering and computer science at MIT and a PhD in computer science at UC Berkeley. Her CV lists a Google research scientist position from 2018 to 2024 with the DeepMind team, formerly Google Brain. At Stanford, she has taught deep reinforcement learning and deep multi-task and meta-learning. Her ICML 2024 talk, “What robots have taught me about machine learning,” was included in Latent Space’s conference recordings.[2][3][1]

Reflecting on her career in 2026, Finn said she had originally planned to enter industry rather than pursue a PhD. She valued doctoral research for teaching her to select problems and handle uncertainty—skills she considers useful at the frontier of AI and in startups. She also emphasized that robotics offers substantial software, hardware, data, and infrastructure work without requiring a PhD, describing the choice between doctoral study and industry as a personal decision.[7]