Walden Hit $1.1B in Six Months. The Diffusion Policy Author List Is 8.
Walden Robotics just raised $300M at $1.1B. The real Large Behavior Models talent pool is 8 paper authors and 132 dual-skill professionals.
On July 15, 2026, Walden Robotics came out of stealth with a $300M seed at a $1.1B post-money, co-led by Toyota and Deviation Capital. Six months of incorporation, robots already on a Toyota plant floor, and a valuation that only makes sense if you understand what the investors actually bought: a paper author list. If you are hiring against Walden, Figure, Skild, Apptronik, or the stealth Apple robotics program, the question is not "how do I find humanoid engineers." It is "how do I find the 8 people who wrote Diffusion Policy, and the ~130 who can meaningfully extend it."
Why Walden's $1.1B seed is really a talent acquisition
Walden is a $1.1B bet on a dozen resumes, not a product. The company incorporated in January 2026 as a Toyota Research Institute spinout under MIT professor Russ Tedrake, formerly TRI's SVP of Large Behavior Models, and its robots have been running in a Toyota North America plant since February 2026. That is under two months from pilot to production, which is only credible because the founding team wrote the underlying methods.
Large Behavior Models (LBMs) are the robotics analog of LLMs: a single policy trained on broad demonstration data that transfers across tasks and improves with practice. The technique that made LBMs tractable is Diffusion Policy, published at RSS 2023 and IJRR 2024 by a group split across Columbia, TRI, and MIT. Tedrake was on that paper. So was Benjamin Burchfiel from TRI, now the technical center of gravity at Walden.
The backer list makes the intent obvious. Alongside Toyota and Deviation Capital, the round drew NVIDIA, Boeing, Samsung Ventures, CoreWeave Ventures, Prologis Ventures, Toyota Invention Partners, Toyota Ventures, AE Ventures, Calibrate, Colle, Shine, NextView, Squarepoint, One Madison, KAS, and Menlo. That is not a syndicate assembled around a product demo. That is a syndicate assembled around a person, and around the two or three papers that person's group produced.
The Diffusion Policy author list is 8 people
The ground-truth "inventor" pool for modern robot imitation learning is 8 names on one paper. Every humanoid company selling an LBM story is, on some axis, downstream of these people. The full list from the 2023/2024 versions of Diffusion Policy (arxiv.org/abs/2303.04137):
- Cheng Chi (Columbia, lead author)
- Zhenjia Xu (Columbia)
- Siyuan Feng (Toyota Research Institute)
- Eric Cousineau (Toyota Research Institute)
- Yilun Du (MIT)
- Benjamin Burchfiel (Toyota Research Institute)
- Russ Tedrake (MIT, now Walden CEO)
- Shuran Song (Columbia, now Stanford, runs the REAL lab)
Two of them are already Walden. One (Song) runs the Stanford lab that produces most of the credible grad-student descendants. The other five are exactly the people every humanoid CEO wants on a call this week. If you are running physical AI recruiting and your Boolean is "Toyota Research Institute alumni" or "robotics + ML," you are searching a pool that is roughly 100x larger than the pool that can actually build this.
What Refolk's index says about the extendable pool
Widen the filter one honest step, from paper authors to people who can extend the method, and the pool is 132 in the US. In Refolk's index of professional profiles, only 132 US professionals list both "Diffusion Policy" and "Imitation Learning" as skills. Drop the Diffusion Policy requirement and expand globally to just "Imitation Learning," and the pool grows to 282. That is the entire addressable market for large behavior models hiring, worldwide.
| Filter | Count | Where it comes from |
|---|---|---|
| Diffusion Policy paper co-authors | 8 | arxiv.org/abs/2303.04137 |
| US professionals with Diffusion Policy AND Imitation Learning | 132 | Refolk's index |
| Global professionals with Imitation Learning | 282 | Refolk's index |
| Diffusion Policy specialists as share of Imitation Learning talent (US) | ~47% | Derived from Refolk's index |
| Walden execs named at launch | 4 | Boston Globe |
| Largest single employer of dual-skill talent | Apple (3) | Refolk's index |
The 47% figure is the one to sit with. Nearly half of everyone who self-identifies as an imitation-learning practitioner has anchored specifically on Diffusion Policy. That is not a normal distribution for a method that is three years old. It means the field has consolidated hard around one technique, and the recruiters who understand that get a much cleaner shortlist than the ones searching for "robotics ML."
Walden's $1.1B seed is investors buying an author list. Every downstream humanoid startup paying market comp for these people is underpricing them.
Apple is the biggest quiet buyer, not Figure
Founders assuming they are only bidding against Figure, Skild, and 1X are wrong. In Refolk's index, Apple is the top single employer of professionals with both Diffusion Policy and Imitation Learning skills, ahead of Samsung Electronics, The Bot Company, Waymo, Motional, and Cohere. Academia (Berkeley, Stanford, UT Austin) supplies the long tail.
The competitive set, if you are hiring:
- Public humanoid pure-plays: Figure AI ($1B+ Series C at $39B in September 2025), Apptronik ($520M extension for Apollo), Skild AI ($1.4B for a robot-agnostic foundation model), 1X, Boston Dynamics (Hyundai-owned, running Atlas deployment plans), NEURA Robotics ($1.4B), LimX, Booster.
- Big Tech stealth programs: Apple robotics, Samsung Electronics, and Waymo's manipulation-adjacent groups. These do not show up in the humanoid press but they show up in Refolk's index as the actual paycheck.
- Adjacent labs: The Bot Company, Cohere's robotics-adjacent research, plus whatever Meta is now standing up post-Reality Labs restructuring.
The mistake is treating this as a humanoid-startup bidding war. The dual-skill pool is only 132 people in the US, and Apple already has three of them. The comp benchmark is not Figure's Series C, it is whatever Apple pays a senior IC with equity refreshers.
Why "TRI alum" is the wrong filter
TRI had hundreds of researchers across autonomy, materials science, and human-interactive driving. The Large Behavior Models work sat in one group under Tedrake and Burchfiel. Filtering LinkedIn for "Toyota Research Institute" surfaces a pool that is roughly 100x larger than the pool that can actually build LBMs.
The signals that separate the ~12 people who matter from the several hundred TRI badge-holders:
- Co-authorship on Diffusion Policy (2023) or the LBM technical reports.
- Direct grad-student lineage from Shuran Song's REAL lab (Stanford, ex-Columbia) or Tedrake's Robot Locomotion Group at MIT CSAIL.
- Author credits at RSS and CoRL in the last three years, specifically on manipulation and imitation learning tracks.
- Membership in TRI's remaining LBM group under Gill Pratt, which is the residual talent Walden could not spin out.
This is the exact filtering problem Refolk was built for. Instead of a Boolean that returns 8,000 TRI alumni or 40,000 "robotics engineers," you ask in plain English for "engineers who co-authored Diffusion Policy or trained under Russ Tedrake or Shuran Song" and get the specific ranked shortlist. GitHub stars on the widely-forked real-stanford/diffusion_policy repo will not surface the inventors, because implementers fork and inventors publish.
The Cambridge geography problem Walden has to solve
Walden is in Cambridge, MA. The dual-skill pool clusters in the Bay Area, Stanford, Pasadena, and Berkeley. Every senior hire is a relocation fight, which is why Tedrake told the Boston Globe "the universities are feeding incredible talent" - he means MIT is his local pipeline, and everyone else is a plane ticket.
Refolk's index shows the diffusion policy robotics engineers pool clustered as follows, at least by primary location:
- SF Bay Area and Stanford: the plurality of the 132 US dual-skill professionals.
- Pasadena (Caltech, JPL adjacencies) and Berkeley.
- Cambridge, MA: a small cluster around MIT CSAIL, mostly students and postdocs, plus the residual TRI Cambridge presence.
- New York (Columbia, Cornell Tech): a secondary cluster, historically tied to Shuran Song's Columbia group before she moved to Stanford.
For Walden, this means the named executive hires so far (Andy Marchese from Amazon Robotics as head of hardware, Joe Romano from Berkshire Grey and Kiva as head of software, Dave Johnson from Draper and Dexai as CPO, per the Boston Globe) are the easy part. They are hardware and product veterans, not LBM researchers. The next 20 hires, the ones that actually build the model, are the ones fighting geography.
What this means if you are hiring against Walden
You have to source from three concentric rings, not from a job posting. The Walden launch and the same-week Skild, Apptronik, and NEURA raises mean the top of the funnel is already gone. Every "senior robotics ML engineer" who is going to answer a recruiter cold email has already answered five.
The three rings, in order:
- The 8 Diffusion Policy authors. Two are Walden. Two are academic PIs (Song at Stanford, Du at MIT). Four are the actual free-agent list. If you are not on a call with them this quarter, you are not competing.
- The 132 US dual-skill professionals in Refolk's index. This is where physical AI recruiting actually happens. Filter by current employer (Apple leads with 3), lab lineage, and paper output in the last 18 months.
- The grad students and postdocs at REAL (Stanford), Robot Locomotion Group (MIT CSAIL), and the residual TRI LBM group under Gill Pratt. These are the Toyota Research Institute alumni who actually did the work, not the badge-holders.
The pattern-matching for founders reading the room: Deviation Capital (Colin Beirne's first publicly disclosed humanoid check), Toyota Ventures, and NVIDIA's corp-dev arm are all signaling that the inventor premium is priced at the seed, not the Series A. If you are raising a humanoid Series A on the same author list Walden bought at seed, your comp bands have to reflect that. Otherwise you are shopping the same 132 people with less money and later timing. Sourcing tools like Refolk help you see the pool before you write the comp bands, not after.
FAQ
Who are the 8 co-authors of the Diffusion Policy paper?
The 2023 RSS and 2024 IJRR versions of Diffusion Policy list eight authors: Cheng Chi (Columbia, lead), Zhenjia Xu (Columbia), Siyuan Feng (TRI), Eric Cousineau (TRI), Yilun Du (MIT), Benjamin Burchfiel (TRI), Russ Tedrake (MIT), and Shuran Song (Columbia, now Stanford). Two of them (Tedrake and Burchfiel-adjacent) are now the technical center of Walden Robotics. The rest are the free-agent list every humanoid company in the current cycle is trying to reach.
Why is "TRI alum" a bad filter for LBM hiring?
Toyota Research Institute employed hundreds of researchers across autonomy, materials, and human-interactive driving. The Large Behavior Models work sat in one group under Tedrake and Burchfiel. Filtering on "Toyota Research Institute alumni" surfaces a pool roughly 100x larger than the pool that can extend LBMs. The right filter is paper co-authorship and lab lineage, not a badge.
How does Walden compare to Figure, Skild, and Apptronik?
Walden raised $300M at $1.1B in July 2026 as a six-month-old TRI spinout with production deployments at a Toyota plant. Figure AI closed a Series C above $1B at $39B in September 2025. Skild AI raised $1.4B for a robot-agnostic foundation model, and Apptronik closed a $520M extension for Apollo. Morgan Stanley projects the humanoid market could exceed $5 trillion by 2050. All four are bidding for the same 132-person US pool.
Who is quietly the biggest employer of Diffusion Policy talent?
Apple. Refolk's index shows Apple leads the top-employer list for US professionals with both Diffusion Policy and Imitation Learning skills, ahead of Samsung Electronics, The Bot Company, Waymo, Motional, and Cohere. Humanoid founders assuming they are only bidding against Figure and Skild are missing the largest single buyer in the pool.