Toyota Research Institute
Last reviewed
Jun 4, 2026
Sources
29 citations
Review status
Source-backed
Revision
v1 · 2,696 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 4, 2026
Sources
29 citations
Review status
Source-backed
Revision
v1 · 2,696 words
Add missing citations, update stale details, or suggest a clearer explanation.
The Toyota Research Institute (TRI) is a research and development subsidiary of Toyota Motor Corporation, based in the United States, that conducts applied and fundamental research in artificial intelligence, robotics, automated driving, and energy and materials science. Toyota announced the institute on November 6, 2015, with an initial commitment of about $1 billion over five years, and TRI began operations in January 2016 with roughly 200 employees. It is headquartered in Los Altos, California, in Silicon Valley, and has long maintained a second major site near the Massachusetts Institute of Technology in Cambridge, Massachusetts, plus a former office in Ann Arbor, Michigan, near the University of Michigan. The institute is led by founding chief executive Gill Pratt, a roboticist and former program manager at the U.S. Defense Advanced Research Projects Agency (DARPA) who ran the DARPA Robotics Challenge.
TRI is best known in the AI and robotics community for its work on robot learning: the team co-developed Diffusion Policy in 2023, coined and championed the term Large Behavior Models (LBMs) for robot manipulation, and in 2025 published a large empirical study of LBMs and partnered with Boston Dynamics to put an LBM on the electric Atlas humanoid. TRI is a distinct organization from Toyota's Japan-based Partner Robot Division, which builds the company's older humanoid and assistance platforms such as the T-HR3, the Human Support Robot, and the CUE basketball robot; for that lineage see the Toyota (robotics) article.
Toyota disclosed plans for the institute on November 6, 2015, at a press event in California, framing it as a way to accelerate the company's capabilities where software and AI meet physical systems. The announced budget was approximately $1 billion over five years, and Toyota separately pledged roughly $50 million over five years to fund joint AI research centers at Stanford University and MIT, work that had been set in motion earlier in 2015 through the Stanford and MIT collaborations. TRI was incorporated as Toyota Research Institute Inc. and opened its headquarters in Silicon Valley, initially described as being in Palo Alto near Stanford, with a second facility near MIT in Cambridge. Operations began in January 2016 with about 200 employees.
Toyota recruited Gill Pratt to lead the new company. Pratt had served as a DARPA program manager from January 2010 through August 2015, where he directed the DARPA Robotics Challenge as well as robotics and neuromorphic computing programs. Before DARPA he was a founding faculty member at the Franklin W. Olin College of Engineering and an associate professor at MIT, where he ran the Leg Laboratory. Pratt earned his bachelor's, master's, and doctoral degrees in electrical engineering and computer science from MIT.
At its launch, Toyota set out three primary goals for TRI: improving vehicle safety by reducing the likelihood of crashes, broadening access to driving for people of varying ages and abilities, and adapting Toyota's mobility technology for indoor and non-driving settings, especially to help older people live independently. The company also said the institute would apply its work more broadly, for example to improve manufacturing efficiency and to accelerate scientific discovery in materials.
| Person | Role | Notes |
|---|---|---|
| Gill Pratt | Chief Executive Officer (founding) | Also a Toyota Motor Corporation Fellow and chief scientist; former DARPA program manager (2010 to 2015) |
| Russ Tedrake | Senior Vice President, Large Behavior Models | Holds the Toyota Professorship at MIT (EECS, Mechanical Engineering, and Aero/Astro); leads TRI's robot-learning research |
Pratt was later named a Toyota Motor Corporation Fellow and serves as a chief scientist advising Toyota's broader research, a sign of how closely TRI's leadership is tied to the parent company's technology strategy. Robotics researcher Russ Tedrake, a longtime MIT professor, leads the institute's Large Behavior Models effort and is one of the most visible public faces of TRI's manipulation research.
TRI sits within a wider set of Toyota software and venture organizations that are frequently confused with it but are legally and operationally separate:
TRI's research often feeds Toyota's production engineering and these sister organizations, but the institute itself remains focused on earlier-stage research rather than shipping production software.
TRI organizes its work around four broad divisions: robotics, automated and assisted driving, energy and materials, and human-centered AI, supported by core machine learning research that cuts across them.
TRI's robotics division concentrates on home and assistive robots, motivated by aging populations in Japan and elsewhere and by the goal of helping people "age in place." Rather than pursuing full autonomy as an end in itself, TRI frames its mission as "intelligence amplification" and "mobility for all," using AI to extend human reach, strength, and independence. The division has explored behavior learning from human demonstration, continuous and shared learning across robots, whole-body tactile sensing, simulation-based testing, and unconventional form factors such as a ceiling-mounted gantry robot that descends to perform household chores.
A signature line of work is soft, whole-body manipulation. TRI built the Soft Bubble Gripper, an inflatable latex gripper containing a depth sensor and camera so a robot can "feel" what it holds, and released the design openly. Building on that, TRI introduced Punyo, a torso-up humanoid research platform whose arms and chest are wrapped in compliant materials and tactile sensors, with air-filled bladders along each arm that can be individually pressurized. The name comes from a Japanese word for something soft, cute, and resilient. Punyo is designed to carry and manipulate bulky objects using its whole body, hugging and bracing loads against its chest the way a person might, rather than relying only on fingertip grasps. The fuller technical history of Punyo, the Soft Bubble Gripper, and Diffusion Policy is covered in the Toyota (robotics) article; this page treats them as part of the institute's research portfolio.
TRI coined the term Large Behavior Model to describe a single neural network, trained on large and varied robot data, that can perform many manipulation tasks and pick up new ones from a handful of demonstrations, analogous to how a large language model generalizes across text. The approach grew out of Diffusion Policy, a 2023 method (developed with Shuran Song's group, then at Columbia University) that represents a robot's visuomotor policy as a conditional denoising diffusion process and won wide attention at the Robotics: Science and Systems conference. TRI positioned Diffusion Policy as a building block toward LBMs and a broader robot foundation model program, situating its work within the wider field of embodied AI.
On July 7, 2025, the institute posted a paper titled "A Careful Examination of Large Behavior Models for Multitask Dexterous Manipulation" (arXiv 2507.05331), credited to the "TRI LBM Team" and more than 80 named authors, with Rares Ambrus listed as the corresponding contact. The study trained LBMs built on a diffusion-transformer architecture that ingests camera images, proprioception, and language prompts and outputs chunks of robot actions. According to the paper and TRI's accompanying materials, the models were trained on roughly 1,700 hours of robot data drawn from internal bimanual teleoperation, simulation, the Universal Manipulation Interface, and the public Open X-Embodiment dataset. The team evaluated performance across 29 tasks using more than 47,000 simulation rollouts and about 1,800 real-world trials on dual-arm stations using Franka Panda FR3 arms.
The headline result, emphasized in TRI's communications and widely reported, is that a single pretrained LBM can learn hundreds of tasks and acquire new skills with markedly less data than training a fresh single-task policy, with TRI citing figures of up to 80 percent less data. The paper itself states the more measured claim that finetuned LBMs are more robust to distribution shift and need only a fraction of the data of single-task baselines to reach the same performance, reporting roughly 3 to 5 times less data in challenging settings. A stated motivation of the work was methodological: to provide a careful, statistically rigorous benchmark for robot foundation models, an area the authors argued had been hard to evaluate honestly.
In October 2024 (announced October 16), TRI and Boston Dynamics formed a research partnership to combine TRI's Large Behavior Models with Boston Dynamics' electric Atlas humanoid, with Russ Tedrake and Boston Dynamics' Scott Kuindersma co-leading the Boston-based effort. On August 20, 2025, the two organizations released results showing a single LBM driving Atlas through long, continuous sequences of whole-body manipulation and locomotion: walking, crouching, lifting, and tasks such as packing, sorting, tying a rope, unfolding a tablecloth, and flipping a barstool, while recovering from interruptions like a researcher sliding a box away mid-task.
The technical novelty was that one language-conditioned, end-to-end policy directly controlled the whole robot, treating the hands and the feet as parts of a single action representation rather than splitting low-level walking and balance from arm manipulation as humanoid stacks usually do. Boston Dynamics described the policy as a roughly 450-million-parameter diffusion transformer trained with a flow-matching objective, running at about 30 Hz and predicting short action chunks, with training data collected by full-body virtual-reality teleoperation. The companies presented the demonstration as a step toward general-purpose humanoids and a way to add new skills quickly through human demonstrations. This work parallels other robot-foundation-model efforts in the field such as Mobile ALOHA.
For self-driving, TRI publicly pursued a deliberately two-track strategy that it branded "Guardian" and "Chauffeur." Chauffeur referred to full autonomy, with the human removed from the driving task either everywhere or within a restricted operational design domain, corresponding to higher SAE automation levels. Guardian, by contrast, was an advanced safety system meant to amplify rather than replace the human driver: the person stays in control, while the automated system runs in parallel, watches for hazards (including signs of distraction or drowsiness), and intervenes only when needed to avoid a crash. TRI argued Guardian could be deployed sooner and save lives in the near term, and that it could in principle backstop either a human driver or another company's autonomous system. The institute demonstrated both modes on test platforms and unveiled research vehicles such as the P4 automated-driving prototype at CES. Much of Toyota's later production-bound automated-driving software work moved to TRI-AD and then Woven, while TRI continued more exploratory driving research.
TRI applies machine learning, robotics, and laboratory automation to accelerate the discovery and design of new materials, particularly for batteries and clean-energy applications. Its Accelerated Materials Design and Discovery program partners with universities and companies to use AI to speed up materials research; in May 2021 TRI committed an additional $36 million to the program over four years, and the institute has said it has put well over $70 million into accelerated materials work overall. In a widely cited 2019 result, researchers from MIT, Stanford, and TRI showed that combining large experimental datasets with machine learning could predict the useful cycle life of lithium-ion batteries early and accurately. TRI has also collaborated with institutions such as Northwestern University on algorithms for rapidly synthesizing and screening candidate materials.
TRI's human-centered AI work brings together researchers from behavioral science, machine learning, and human-computer interaction to model human behavior and build AI that augments human decision-making rather than supplanting it. This division reflects the institute's overarching philosophy, shared across its robotics and driving work, that technology should keep people engaged and in control, a stance Toyota frequently summarizes as "mobility for all."
TRI is one of the more academically engaged corporate AI labs, publishing openly, releasing datasets and hardware designs, and seeding a venture fund, while remaining tied to a large automaker's long-term priorities around safety, aging, and clean energy. Its robot-learning research, especially Diffusion Policy and the Large Behavior Models program, has been influential in the shift toward learned, generalist robot policies, and its 2025 collaboration with Boston Dynamics is among the most visible demonstrations of a single learned model controlling a full humanoid for whole-body manipulation.