Skip to content
No PriorsNo Priors

No Priors Ep. 107 | With Physical Intelligence Co-Founder Chelsea Finn

This week on No Priors, Elad speaks with Chelsea Finn, cofounder of Physical Intelligence and currently Associate Professor at Stanford, leading the Intelligence through Learning and Interaction Lab. They dive into how robots learn, the challenges of training AI models for the physical world, and the importance of diverse data in reaching generalizable intelligence. Chelsea explains the evolving landscape of open-source vs. closed-source robotics and where AI models are likely to have the biggest impact first. They also compare the development of robotics to self-driving cars, explore the future of humanoid and non-humanoid robots, and discuss what’s still missing for AI to function effectively in the real world. If you’re curious about the next phase of AI beyond the digital space, this episode is a must-listen. Sign up for new podcasts every week. Email feedback to show@no-priors.com Follow us on Twitter: @NoPriorsPod | @Saranormous | @EladGil | @ChelseaFinn Show Notes: 0:00 Introduction 0:31 Chelsea’s background in robotics 3:10 Physical Intelligence 5:13 Defining their approach and model architecture 7:39 Reaching generalizability and diversifying robot data 9:46 Open source vs. closed source 12:32 Where will PI’s models integrate first? 14:34 Humanoid as a form factor 16:28 Embodied intelligence 17:36 Key turning points in robotics progress 20:05 Hierarchical interactive robot and decision making 22:21 Choosing data inputs 26:25 Self driving vs robotics market 28:37 Advice to robotics founders 29:24 Observational data and data generation 31:57 Future robotic forms

Sarah GuohostChelsea Finnguest
Mar 19, 202535mWatch on YouTube ↗

At a glance

WHAT IT’S REALLY ABOUT

Chelsea Finn on building general-purpose robots through data and hierarchy

  1. Chelsea Finn, Stanford professor and co-founder of Physical Intelligence (PI), discusses her decade-long journey in robotics and her company’s mission to build general-purpose AI models that can control many different robots in the physical world.
  2. PI focuses on scaling diverse real-world robot data, sharing hardware designs, and developing foundation models that generalize across tasks, environments, and embodiments rather than optimizing for a single narrow application.
  3. Finn explains their architecture combining transformers with pretrained vision-language models, the importance of teleoperated data collection, and their new hierarchical interactive robot system that decomposes long-horizon tasks and incorporates natural language interaction.
  4. She reflects on challenges in robotics versus self-driving, the underrated complexity of motor control, likely form factors for future robots, and why openness and community-building may be essential for the field to succeed at all.

IDEAS WORTH REMEMBERING

5 ideas

Generalist robot models require massive, diverse real-world data, not just more compute.

Finn emphasizes that the primary bottleneck is breadth and diversity of robot experience—many buildings, objects, and tasks—rather than simply scaling model size or FLOPs.

Teleoperated, robot-specific data is irreplaceable, even when leveraging human and internet videos.

While web and human demonstration data help with concepts and semantics, robots must still practice using their own bodies to acquire low-level motor competence, analogous to how humans learn physical skills.

Cross-embodiment learning can recycle data and accelerate progress across robot platforms.

PI’s research shows that policies trained on pooled data from many different robots can outperform per-lab models, and that changing hardware no longer forces you to discard past datasets.

Hierarchical and language-interactive control is key for long, complex tasks.

Their HI robot system separates high-level step selection (conditioned on natural language prompts) from low-level motor control, enabling tasks like sandwich-making with on-the-fly user corrections.

Openness—sharing code, weights, and even hardware designs—is a deliberate strategic bet.

PI believes the bigger risk is that no one solves general-purpose robotics; building a robust ecosystem, attracting top researchers, and seeding better hardware outweighs IP-protection concerns at this phase.

WORDS WORTH SAVING

5 quotes

We're trying to build a big neural network model that could ultimately control any robot to do anything in any scenario.

— Chelsea Finn

The number one thing is just getting more diverse robot data.

— Chelsea Finn

I think the biggest risk with this bet is that it won't work. I'm not really worried about competitors; I'm more worried that no one will solve the problem.

— Chelsea Finn

People underestimate how much intelligence goes into motor control.

— Chelsea Finn

I would love if we could give our robots skin.

— Chelsea Finn

Chelsea Finn’s background and evolution of neural-network-based robot controlPhysical Intelligence’s mission: generalist foundation models for many robot platformsData strategy: large-scale teleoperation, diverse environments, and web/vision-language pretrainingArchitecture and hierarchy: high-level planning plus low-level motor control policiesOpen vs. proprietary approaches in robotics research and commercializationTechnical and practical challenges in general-purpose robotics versus self-drivingFuture of robot form factors, sensors, and the role of embodied intelligence

High quality AI-generated summary created from speaker-labeled transcript.

Get more out of YouTube videos.

High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.

Add to Chrome