HomeEventsMinds in Motion at MIT, Physical AI SummitWhat Robots Can Learn, and How Fast
Keynote · 10:00 · Models Scale

What Robots Can Learn, and How Fast

Minds in Motion at MIT, Physical AI Summit · 30 Oct 2026 · Cambridge, United States

About this session

Russ Tedrake, MIT professor and Co-Founder and CEO of Walden Robotics, on how robots learn new tasks, what they can do on a factory floor today, and what is still out of reach.

On stage

moderatorGreg HuangSenior Editor, The Boston Globe
speakerRuss TedrakeCEO & Cofounder, Walden Robotics

Companies on stage

Russ Tedrake, Professor at MIT CSAIL; CEO & Cofounder of Walden

Walden Robotics builds and deploys general-purpose wheeled robots with arms for manufacturing and logistics customers.

Signals from this session

5:01 in the recordingRuss TedrakeRobots demonstrated peanut-butter spreading and shoelace tying, but high reliability remains unresolved. Robotic systems achieved tasks including spreading peanut butter on toast and tying shoelaces. The capability is present at roughly 60% performance, while deployment would require approximately 99.9999% (“five nines”) reliability; the systems can also be programmed rapidly.
5:01 in the recordingRuss TedrakeRobotic manipulation systems can now tie shoelaces and be rapidly programmed from shared code. The remaining challenge is achieving highly reliable operation rather than demonstrating the task at approximately 60% success.
5:01 in the recordingRuss TedrakeRuss Tedrake's team found a way to apply image diffusion models to robot manipulation. Tedrake said his team found the key ingredient for applying diffusion models built for images to robot arms, and that a mature pipeline let them make it work at high quality.

Members see another 6 signals from this session. To read them, subscribe or sign in.

Viewpoints from this session

0:18 in the recordingRuss TedrakeSpring Flamingo was the exciting new frontier in robotics and legged robots at MIT around 1999–2000.
1:00 in the recordingRuss TedrakeThe DARPA Robotics Challenge was a formative moment for the robotics field.
1:24 in the recordingRuss TedrakeGenerative AI makes it possible to program robots through human imitation for tasks that would previously have been very difficult to model, much like GPT applied to robotics.

Members see another 55 viewpoints from this session. To read them, subscribe or sign in.

Other sessions

09:30FiresideThe World's Largest Robot FleetTye Brady, Amazon Robotics · moderated by Aaron Pressman, The Boston Globe09:30PanelFuel for Physical AIMichael Healy, LabyrinthAI · Chase Olle, Robot on Rails · Enja Ahearn, AWS · Natalie Matushevsky, AWS (AI Models, Silicon and Quantum) · Romeo Catracchia, AutonomUS · moderated by Aaron Edell, AWS09:30PanelTrained and TestedBerk Calli, Robotics at Worcester Poly Tech · Raul Romero, Kite ML · Kentaro Barhydt, MIT MechE · moderated by Rachael Harkavy, Link Ventures10:00PanelShow, Don't CodeMichael Healy, LabyrinthAI · Robin Walters, Braid Robotics · Gabriela Torres V., Massachusetts AI Hub · Anshul Mittal, Oracle AI · moderated by Oded Maron, MassRobotics10:00PanelSoftware for the Factory FloorJosef Waltl, Software Defined Automation · Peter Howard, Realtime Robotics · Alex Tilita, MatterWave · Pavan Pant, Tristar AI · moderated by Tyler Bouchard, Flexxbotics10:00TalkRobotic Blood DrawRavi Kaushik, ROPHAI

Full program →

Members read every signal and every company file.

They also get the Frontier Tech Digest every morning and can ask their own questions. What Frontier Tech Intelligence is →

Subscribe · $99 a month
Independent coverage, not affiliated with any organizer.
↑