Walden Robotics Raised $300M. The Real LBM Author Pool Is 58.

On July 15, 2026, Walden Robotics stepped out of stealth with $300M at a $1.1B valuation, a Toyota Research Institute spinout led by MIT's Russ Tedrake, with robots already running at a Toyota plant in North America since February 2026. Every recruiter I know is now typing "robotics ML engineer" into LinkedIn and getting back a five-figure haystack. The engineers who have actually shipped a Large Behavior Model live on a single arXiv author list, and that list has 58 names.

Why the Walden Robotics hiring race is really a fight over one arXiv paper

Walden Robotics hiring is best modeled as a bidding war over the 58 named authors on arXiv:2507.05331, not as a generic "physical AI recruiting" pull. The company's product is explicitly built on Diffusion Policy and Large Behavior Models, which Businesswire quoted as "the frontier model class that powers Walden's robots." That phrasing narrows the credible candidate pool from the broad "robot learning" market down to people who have written LBM code that runs on real hardware.

The July 2025 TRI paper, "A Careful Examination of Large Behavior Models for Multitask Dexterous Manipulation," is the reference implementation of that stack. It extends the Diffusion Policy paradigm across simulated and real robot data. If you can name every human who contributed to that paper, you can name the shortlist for every LBM-native role Walden, Boston Dynamics, Figure, Apptronik, 1X, Physical Intelligence, Skild AI, Nvidia GEAR, and DeepMind Robotics are currently trying to fill.

58

What "Large Behavior Models engineers" actually means

A Large Behavior Model is a multitask robot manipulation policy trained across simulated and real robot data using the Diffusion Policy paradigm. Diffusion Policy, originally from Cheng Chi and Shuran Song's group at Columbia, is a class of visuomotor policies that generates robot actions by iteratively denoising a trajectory. In practice, "Large Behavior Models engineers" today means people who have touched three specific things:

• Policy architecture and training code for diffusion-based manipulation models.
• A simulation stack, in most cases Drake, wired to the LBM evaluation harness.
• Real robot data collection and sim-to-real transfer for dexterous manipulation.

That is a much narrower spec than "robotics ML." It excludes locomotion specialists, classical motion planners, perception-only researchers, and most of the RL crowd. It maps almost one-to-one onto the TRI LBM Team author list plus the Boston Dynamics contributors on the August 2025 Atlas LBM collaboration (Scott Kuindersma, Ben Burchfiel, Siyuan Feng, Kerri Fetzer-Borelli, Lucas Manuelli, Pat Marion, Eric Cousineau, and others).

The high-signal first-author names are worth memorizing if you are sourcing this space: Jose Barreiros, Andrew Beaulieu, Aditya Bhat, Rick Cory, Eric Cousineau, Hongkai Dai, Ching-Hsin Fang, Kunimatsu Hashimoto, Muhammad Zubair Irshad, Masha Itkina, Naveen Kuppuswamy, Kuan-Hui Lee, Katherine Liu, Dale McConachie, Ian McMahon, Haruki Nishimura, Calder Phillips-Grafflin, Charles Richter, Paarth Shah, Krishnan Srinivasan, Blake Wulfe, Chen Xu, Mengchao Zhang.

The numbers: 1,007 fuzzy matches vs. 4 verified skill holders

The supply picture, pulled from Refolk's index against the arXiv author roster, is brutally lopsided.

Read the last row twice. The paper is a bigger sourcing list than the platform skill tag. That is a strange sentence, and it is true. The people who built Diffusion Policy do not tag themselves with it, because when you invent a technique you do not put it in your Skills section. You put "robot learning," "imitation learning," or a lab affiliation and move on.

The paper is a bigger sourcing list than the platform skill tag. Source the paper.

Why Diffusion Policy researchers refuse to self-label

Diffusion Policy researchers self-label as generalists because the term was coined in 2023 and its authors think of it as one tool among several. Cheng Chi and Shuran Song at Columbia published the original work; Toyota Research Institute extended it into LBMs. Neither group treats "Diffusion Policy" the way a React developer treats "React." It is a technique, not an identity.

The mechanism is straightforward:

1. Researchers optimize their profiles for lab and advisor prestige (MIT CSAIL, TRI, Columbia AI Robotics, Stanford), not for technique keywords.
2. Skills sections on LinkedIn cap out at a handful of tags. Nobody wastes one on a specific paper's method name.
3. The community reads arXiv and GitHub, so credibility is signaled through commits to columbia-ai-robotics/diffusion_policy or ToyotaResearchInstitute/lbm_eval, not through profile tags.

This is why keyword sourcing collapses here. If you type "Diffusion Policy" into LinkedIn you get 4 US profiles. If you reverse the lbm_eval contributor graph on GitHub, cross it against the arXiv author list, and enrich with employer and location, you get the actual pool. That reversal is the exact gap Refolk closes: describe the person in plain English ("US-based ML engineer who has contributed to the TRI LBM eval repo or co-authored the Diffusion Policy paper") and get a ranked shortlist across GitHub, LinkedIn, and the open web.

The TRI spinout has already claimed the primary contributors

The Toyota Research Institute spinout dynamics mean Walden gets first-refusal on most of the LBM paper's primary contributors, and Boston Dynamics has already locked its own slice through the Atlas LBM collaboration. That leaves competitors fighting over a residual pool of roughly 20 to 30 second authors.

Who is already spoken for

• Walden Robotics. Tedrake was SVP of Large Behavior Models at TRI. Any TRI LBM first author who followed him out is already an employee or an easy call.
• Boston Dynamics. The August 2025 BD + TRI Atlas paper puts Scott Kuindersma (BD Lead), Ben Burchfiel, Siyuan Feng, Kerri Fetzer-Borelli, Lucas Manuelli, Pat Marion, and Eric Cousineau on record as BD-adjacent LBM talent.
• Toyota Research Institute itself. TRI is not dissolving. It retains a working LBM team that will resist poaching, especially given Toyota is a co-lead investor in Walden alongside Deviation Capital, with Nvidia, Boeing, Samsung Ventures, CoreWeave Ventures, and Prologis Ventures on the round.

Who is actually available

Everyone else is competing for what remains: second-author contributors, the lbm_eval external contributor graph, Columbia's original Diffusion Policy graduates, MIT Robot Locomotion Group alumni, and CoRL and RSS 2024 and 2025 author lists that overlap with manipulation and imitation learning. This is a hundred-ish-person pool globally, not a thousand-person pool.

Wheels, not legs: the hiring spec Walden actually needs

Walden's decision to ship wheeled robots first means it does not need bipedal locomotion talent, which is the scarcest and most contested humanoid sub-pool. Tedrake told Bloomberg that factory workers "aren't ready, and they don't want" walking humanoids yet. What Walden needs instead is manipulation-policy and sim-to-real engineers, which is exactly what the LBM first-author list is optimized for.

This is a real edge for the recruiting team that catches it. Consider the delta:

• Bipedal locomotion pool. Dominated by Boston Dynamics and Agility Robotics alumni. Figure, 1X, Apptronik, and Tesla Optimus are all bidding on the same roughly 100 people.
• Manipulation LBM pool. Dominated by TRI, Columbia, MIT, and Stanford graduates. Walden effectively owns the on-ramp.

If you are hiring against Walden without understanding this split, you will overpay for bipedal specialists Walden does not want and underbid on the manipulation people it will pay any number for. Describing the split in plain English inside a sourcing tool ("manipulation policy researchers with Drake simulation experience, not bipedal locomotion") is faster than building two separate Boolean strings and deduping them by hand.

Where to source when the arXiv list runs out

When you have contacted all 58 named authors, source adjacent artifacts before you fall back to keyword searches. The high-signal artifacts are:

• ToyotaResearchInstitute/lbm_eval GitHub contributor graph. The 49-task Drake-based sim benchmark is a de facto credential. External contributors are self-selected LBM enthusiasts.
• columbia-ai-robotics/diffusion_policy repo. The original codebase. Its issue and PR history names people who ran Diffusion Policy on their own hardware.
• CoRL 2024 and 2025 author lists. Conference on Robot Learning is the venue where manipulation policy work lands first.
• RSS 2024 and 2025 author lists. Robotics: Science and Systems, especially the imitation learning and manipulation sessions.
• MIT Robot Locomotion Group alumni. Tedrake's group. Direct pipeline into Walden's engineering culture.
• Shuran Song's lab alumni (now at Stanford). Diffusion Policy's origin lab.
• TRI's public LBM page at tri.global/our-work/large-behavior-models, which occasionally names collaborators the paper does not.

252x

The Boston geographic squeeze nobody is pricing in

Three of the top LBM employers - TRI, Walden, and Boston Dynamics - all sit within 15 miles of each other in the Cambridge-Waltham corridor, which means any LBM author unwilling to relocate becomes a three-way bidding war on home turf. The Boston Globe already flagged Walden's round as one of the largest robotics deals ever in the Boston ecosystem, trailing only Motional per PitchBook.

The implication for physical AI recruiting is uncomfortable for California and Texas shops:

• Figure (California) and Apptronik (Texas) can pay top-of-market, but they cannot offer a 15-minute commute to a former TRI colleague's new startup.
• Relocation packages get eaten by dual-career households and Cambridge school districts.
• The people most likely to move for Figure or Apptronik are second authors and adjacent graduates, not the first-author core.

If you are running physical AI recruiting from outside the Boston corridor, price this in early. Your realistic pool is smaller than the arXiv list suggests, because roughly half of it is functionally non-relocatable. Morgan Stanley projects the humanoid market could top $5 trillion by 2050, and that ceiling is what forces every non-Boston shop to keep bidding anyway.

The takeaway for anyone hiring against Walden

Stop searching for "robotics ML engineer" and start searching for the paper. The 1,007 headline matches are noise. The 58 named authors are the market. The Diffusion Policy skill tag is a trap. The GitHub contributor graph on lbm_eval and diffusion_policy is the real signal, and it maps to fewer than 200 humans worldwide.

Walden Robotics did not just raise $300M. It bought first pick of a nameable list, and every other humanoid company is now recruiting against a residual that fits on one page. If your sourcing tool cannot ingest an arXiv author list and enrich it against GitHub, LinkedIn, and the open web in one query, you are hiring blind.

FAQ

How many people are actually qualified to build a Large Behavior Model in production?

The globally credible pool is roughly 58 named authors on arXiv:2507.05331, plus the Boston Dynamics contributors on the August 2025 Atlas LBM collaboration, plus a smaller set of Columbia and Stanford graduates from the original Diffusion Policy work. Fewer than 200 humans have shipped this stack. In the US, only 4 profiles explicitly list Diffusion Policy as a skill, so relying on LinkedIn skill tags will miss the vast majority of qualified people.

Why is a keyword search for "Diffusion Policy" almost useless?

Because engineers who built the technique do not tag themselves with the technique's name. Diffusion Policy was coined in 2023 by Cheng Chi and Shuran Song's group at Columbia, and the community treats it as one tool among several. Practitioners self-label as "imitation learning," "robot learning," or by lab affiliation (MIT CSAIL, TRI, Columbia AI Robotics). The real credential is a GitHub commit history on columbia-ai-robotics/diffusion_policy or ToyotaResearchInstitute/lbm_eval, not a LinkedIn Skills entry.

What is different about Walden's hiring spec compared to Figure or Apptronik?

Walden ships wheeled robots first, per Tedrake's Bloomberg interview, so it does not compete for bipedal locomotion talent. It hires manipulation-policy and sim-to-real engineers, which map almost one-to-one onto the TRI LBM first-author list. Figure, Apptronik, 1X, and Tesla Optimus are largely still competing for bipedal specialists, a different and even scarcer pool dominated by Boston Dynamics and Agility Robotics alumni.

If most of the arXiv authors are already at Walden or Boston Dynamics, where should I source next?

Work outward from the paper. Start with the ToyotaResearchInstitute/lbm_eval contributor graph and the columbia-ai-robotics/diffusion_policy repo. Add CoRL and RSS 2024 and 2025 author lists filtered to manipulation and imitation learning sessions. Add MIT Robot Locomotion Group alumni and Shuran Song lab alumni. That gives you a realistic secondary pool of maybe 100 to 150 people worldwide, small enough to run a fully personalized outbound campaign against.