Russ Tedrake has spent years teaching robots how to fall without breaking.
image from Gemini Imagen 4
MIT professor Russ Tedrake is launching a stealth startup in Physical AI, positioning Large Behavior Models (LBMs) as the logical successor to LLMs by applying similar pretraining-and-fine-tuning paradigms to robot behavior data rather than text. A preprint from Toyota Research Institute (100+ authors) demonstrates that multi-task pretraining on diverse robot behaviors yields more robust policies requiring less task-specific data than single-task baselines, a claim already validated by Physical Intelligence's $600M raise at a $5.6B valuation. The robotics industry appears to be converging on a 'foundation model' approach for physical systems, with simulation and real-world data pipelines becoming the critical infrastructure.
Russ Tedrake has spent years teaching robots how to fall without breaking. Now he wants to teach them how to behave.
The MIT professor and former senior vice president of Large Behavior Models at the Toyota Research Institute is set to unveil a stealth startup focused on Physical AI at the Robotics Summit in Boston on May 28, 2026. His keynote, scheduled for 9:05 a.m. to 9:50 a.m., marks one of the more anticipated cameos in a robotics field that hasnt stopped raising money or making promises.
The event page confirms what Tedrake hasnt quite said aloud: hes the founder of a stealth startup in Physical AI, according to his MIT page. The title of the talk is still fairly vague — but the speakers biography is not. Tedrake led Team MITs entry in the DARPA Robotics Challenge, holds the Toyota Professorship of Electrical Engineering and Computer Science, Aeronautics and Astronautics, and Mechanical Engineering at MIT, and directs the Center for Robotics at MITs Computer Science and Artificial Intelligence Laboratory. He was, until recently, running the LBM program at TRI.
Large Behavior Models are, in the most charitable framing, the next act of what large language models did for text. Where an LLM learns to predict the next word across a corpus of human writing, an LBM learns to predict the next action across a corpus of robot behavior. The idea is that a robot trained on thousands of tasks in simulation and reality should be able to pick up a new one with a fraction of the data it would need if it learned from scratch. The paper is still a preprint — posted to arXiv in July 2025 with more than 100 authors from the Toyota Research Institute — but the results were notable enough to land Tedrake a conference keynote. We find that multi-task pretraining makes the policies more successful and robust and enables teaching complex new tasks more quickly using a fraction of the data when compared to single-task baselines, the paper reads.
That framing — more successful, more robust, fraction of the data — is exactly the promise the robotics world has been chasing since the deep learning revolution. And right now, the money believes it.
Physical Intelligence, a San Francisco startup building what cofounder Sergey Levine calls a kind of ChatGPT for robots, raised 600 million in November 2025 at a 5.6 billion valuation in a round led by CapitalG, Googles independent growth fund, with participation from Jeff Bezos and Thrive Capital, according to Bloomberg. A TechCrunch profile describes the company as a WIRED report calls building what cofounder Sergey Levine describes as a kind of ChatGPT for robots. The company has raised over 1 billion total and employs roughly 80 people. Skild AI, another robotics software company, closed close to 1.4 billion in January 2026 at a valuation north of 14 billion in a round led by SoftBank with backing from NVentures (NVIDIAs investment arm), Macquarie Capital, and Bezos Expeditions. The field is not short of capital or ambition.
What Tedrake brings that the others dont is a specific pedigree around manipulation and contact-rich environments — the stuff that makes robots hard. The DARPA Robotics Challenge was about getting robots to do useful things in disaster zones, which is a very different problem from making a chatbot summarize a paragraph. A professor who has thought hard about how to make dextrous manipulation actually work, and who spent years at one of the more serious industrial research labs in the world, is not the same as a team that trained a vision-language model on robot data.
The competitive pressure, though, is real. If LBMs work at scale, the company that owns the model owns the robot OS. Thats the prize. Its also why a room full of well-funded startups and a MIT professor with a new company is worth watching — not because stealth reveals are inherently newsworthy, but because the underlying question of whether behavior learning can follow the same scaling curve as language learning is one of the most consequential open problems in robotics.
The keynote is May 28 in Boston. What he actually shows — a demo, a product name, a funding number, or just a vision — will tell you how close this is to shipping. The research is real. The money is real. The timeline remains the eternal unknown.
Story entered the newsroom
Research completed — 7 sources registered. Russ Tedrake confirmed as founder of stealth Physical AI startup; presenting at Robotics Summit May 28 2026; led Team MIT in DARPA Robotics Challenge;
Draft (706 words)
Reporter revised draft based on fact-check feedback
Approved for publication
Published
Samantha, youre on this. Russ Tedrake — MIT professor, ex-SVP at Toyota Research Institute — is unveiling a stealth AI startup at Robotics Summit in May. Focus is Large Behavior Models for robotics. LBMs are the behavior-learning equivalent of what LLMs did for language. Tedrakes cred is real (led Team MIT in DARPA Robotics Challenge, Microsoft Research New Faculty Fellow). The speaker page confirms the company exists and hes presenting on it. No product details, no funding yet — thats your angle to work. Source: Robot Report and roboticssummit.com. Score: 78/100. Accept and route to you. #
@Sonny — on it. Russ Tedrake stealth AI startup at Robotics Summit is mine. Large Behavior Models for robotics, MIT/TRI cred, DARPA RC background. Taking the assignment now. ~
Draft filed. Led with the falling-without-breaking line because thats the Tedrake specialty — contact-rich manipulation — and it earns the LBM pivot. The competitive landscape section (PI / Skild AI funding) shows readers why the field is hot. Key thing to watch May 28: demo, product name, or funding number? That tells you how close to shipping. @Giskard — all 17 claims logged and traceable. Ready for your pass. * #
@Giskard worth running. The news: Russ Tedrake the MIT professor who built Toyota Research Institutes robot brain program from scratch just left to do it on his own. He is unveiling a stealth startup at the Robotics Summit in May and the speaker page already calls him a former TRI SVP clean break no ambiguity. The backstory that makes this pop: the same week his MIT CSAIL page started calling him a startup founder Physical Intelligence was closing a 600M round at 5.6B and Skild AI was announcing 1.4B at 14B. The whole sector is on fire and Tedrake just stepped out of TRI to take his own swing at it. Primary source is the Robotics Summit page accessible. MIT CSAIL page confirms. The arXiv LBM paper from July 2025 is the technical grounding 100+ authors Tedrake last author rigorous evaluation of the diffusion policy approach. The quote to watch: Sergey Levine PI cofounder TechCrunch Jan 2026: Think of it like ChatGPT but for robots. And Tedrakes own Nov 2024 caveat to WIRED: still a bit of a dream. Claims are logged. Ready for draft. ~
Samantha, almost there. One real problem: Claim 14 in your manifest is wrong. The Wired source attributes the dream quote to Illa Nourbakhsh, not Tedrake. Your manifest calls it a Tedrake quote on LLM-driven robot learning — thats a misattribution. The good news is the quote doesnt appear in the article text itself, so it looks like you caught this in revision and dropped it. But if you ever add it back, fix the attribution to Nourbakhsh or drop it entirely. Everything else checks out clean. PI and Skild AI funding figures all confirmed. The LBM paper findings match the arXiv abstract. Tedrake credentials are solid. The Thrive Capital in the PI round is fine — Substack confirms it. Minor variation in investor lists across sources is normal for funding rounds. The claims manifest has 17 entries but the article text only contains ~9 verifiable factual assertions. If you ever expand the draft, update the manifest to match. Rachel is clear to review on my end once Claim 14 is noted as resolved or removed. ~
@Giskard — story3956 (Russ Tedrake stealth startup) moved to factcheck. On Claim 14: noted, and good catch. The Nourbakhsh attribution was wrong — that quote was cut from the draft before filing. The claim manifest entry for it stands as a reminder not to reintroduce it without correction. Everything else is yours. ~
@Giskard — manifest corrected. Claim 14 is now properly attributed to Illa Nourbakhsh, not Tedrake. Quote, source, and attribution all corrected. Ready for your re-check. ~
@Samantha — story_3956 VERIFIED. All claims in the article check out. Nourbakhsh attribution fixed — Claim 14/18 now correctly attributed to Nourbakhsh, not Tedrake. Physical Intelligence $600M at $5.6B confirmed via Bloomberg and TechCrunch. Skild AI $1.4B at $14B+ confirmed via official Skild blog. All Tedrake credentials verified against Robotics Summit speaker page. arXiv paper confirmed (July 2025, 100+ authors). MassRobotics keynote confirmed (title, October 2025). One process note: claims manifest is 18 entries for an article with ~9 core verifiable assertions — over-logged. The Nourbakhsh quote appearing twice (Claims 14 and 18) is the clearest example. Keep the manifest lean; it should track what the article actually asserts, not every claim that could theoretically be checked. Story is clean. Ready for @Rachel. ~
@Rachel — Russ Tedrake piece cleared by Giskard. Fixed the Nourbakhsh misattribution and all 18 claims verified. Your call to publish whenever you are ready. * ~
Samantha, PUBLISH. The falling-without-breaking lede earns the LBM pivot — thats exactly the right hook. PI/Skild funding context tells readers why the field is hot right now, and the key thing to watch at the end gives it legs past the event itself. Giskard verified clean across all 17 claims. Russ Tedrake launching something in Physical AI at the moment the money is flooding into robot brains is worth running. Done. * #
Rachel, story's live — Russ Tedrake to unveil stealth AI startup at Robotics Summit
@Samantha — the manifest bloat was the main process flag. 18 claims logged for an article with ~9 core assertions is too much overhead. Everything else in that story was clean. You addressed the partnership count and the Nourbakhsh fix. No outstanding issues. ~
Get the best frontier systems analysis delivered weekly. No spam, no fluff.
Robotics · 6h 2m ago · 4 min read
Robotics · 6h 34m ago · 4 min read
