Walden Robotics, born out of the Toyota Research Institute (TRI), has begun moving research-robot intelligence into real factory work. On July 15, 2026, the company emerged from stealth with a $300 million seed round and an $1.1 billion valuation. Having spun off from TRI in January, the company says it has been running production operations at Toyota plants in North America since February. According to Walden, the move from initial trials to actual operations took less than two months. Walden's aim lies in linking a research team, capital, and a factory into a single learning loop.
That said, it would be premature to read the phrase "getting smarter while working" as meaning the robot autonomously rewrites itself during operation. What TRI's peer-reviewed paper demonstrated is that fine-tuning individual tasks after large-scale pretraining can reduce the amount of demonstration data required. How Walden's commercial units actually collect failures on the factory floor, and who approves any retraining, has not been disclosed.
$300 Million and a Factory Transition in Under Two Months
The funding round was co-led by Toyota Motor Corporation and Deviation Capital. On the Toyota side, the parent company, Toyota Invention Partners, and Toyota Ventures all participated. Other participants included NVIDIA and CoreWeave Ventures, along with Boeing, Samsung Ventures, and Prologis Ventures. Companies that supply computing infrastructure and companies involved in manufacturing and logistics joined the same round.
Walden handles everything in-house, from hardware to AI, deployment software, and field applications. Its initial focus is manufacturing and logistics. The company's official site lists four categories of work: loading and unloading parts at machine tools, tool setup, parts kitting, and assembly. Rather than a conventional industrial robot that rapidly repeats a single fixed task, Walden aims to build a machine that can adapt to process changes and product variation.
Toyota is both an investor and the first major user. According to Walden, Toyota plants in North America have been running production work since February. The transition from initial trials to actual operations took less than two months. Even so, the plant names and the number of units deployed have not been disclosed. Which processes have been entrusted to the robots, how many units they process per hour, and how many times humans had to intervene also remain unknown. The company says the move to real operations marks a step beyond a research demo, but the numbers needed to gauge its maturity as mass-production equipment are yet to come.
What Did the Large Behavior Model Learn?
The technical starting point for Walden lies in Diffusion Policy, developed by TRI and university researchers. The system takes in camera images and tactile information and, much like how image-generation AI refines an image out of noise, progressively builds a candidate sequence of movements. After the robot executes part of that sequence, it observes the changed surroundings and recalculates the next sequence of movements. Its strength lies in being able to learn, from human demonstrations, how to handle objects such as cloth, liquids, or reflective surfaces—things whose shape and motion are hard to define mathematically.
In 2023, TRI announced that it had trained robots on more than 60 kinds of dexterous tasks using a method in which humans taught movements via a haptic device while attaching goals described in language. New motions could be rolled out from just a few dozen demonstrations. Rather than rewriting the program for each new process, the idea is to expand the robot's skills by adding data.
Expanding this approach to a large number of tasks produced the Large Behavior Model (LBM). A TRI paper published in Science Robotics in 2026 trained multiple LBMs using roughly 1,700 hours of demonstration data. Evaluation involved over 1,800 trials on real hardware and more than 47,000 trials in simulation. The physical trials were conducted on a tabletop rig equipped with two Franka FR3 arms, with 50 runs performed for each condition.
The results were clear. Fine-tuning a pretrained LBM for an individual task reduced the amount of data needed to match the performance of a model trained from scratch by a factor of three to five. The improvement was even larger under conditions where object placement and the environment differed from those used during training. What Walden is trying to bring into the factory is precisely this property—reducing the data required each time a new process is taught.
What "Learning on the Job" Actually Involves Remains Unclear
The same paper also notes limitations that tend to get dropped from promotional material. An LBM without task-specific fine-tuning did not consistently outperform a model trained from scratch on a single task. The benefit appears only after a pretrained model has been adapted to the actual task at hand. Furthermore, the variance from retraining runs was not included in the evaluation, and the paper notes that small improvements could be masked by noise in physical trials.
One conceivable model for continual learning in operation would be a loop in which failures and human interventions collected during operation feed back into retraining, which then goes through safety checks before returning to the factory floor. However, Walden has not disclosed either the frequency of retraining or the safety-check procedures involved. Nor does TRI's experimental results guarantee online learning—automatically updating a model's parameters while it's running. In a factory, a single malfunction can halt equipment or cause quality defects, so learning speed and change management cannot be treated as separate concerns.
Walden's careers page hints at the verification framework supporting its operations. The company is hiring for post-training via reinforcement learning and data design, as well as separate roles focused on policy evaluation and experimental infrastructure. On the robotics side, it is hiring for functional safety, reliability testing, and release management. Rather than sending a single model into the factory, Walden appears to be building an organization around data and verification as part of the product itself.
Publicly available information about the commercial units lacks per-process cycle times and uptime rates. Intervention counts and recovery times are also not disclosed, nor are safety certifications or deployment pricing. The result confirmed on a research tabletop rig—reducing required data by a factor of three to five—cannot be directly translated into labor cost savings or an investment payback period.
Toyota Isn't Betting on One Company
At its earnings briefing in May 2026, Toyota cited as its strength the fact that its factories worldwide produce roughly 10 million vehicles a year and that skilled workers can train robots. Parts transport and parts picking are already being piloted, and Toyota has indicated plans to eventually expand beyond its own factories. The idea is to use the Toyota Production System—which accumulates large volumes of real-world work and continuous improvement—as a training environment for robots.
Walden is a key candidate in this effort, but it is not the only approach Toyota has chosen. On February 19, 2026, Toyota Motor Manufacturing Canada signed a Robot-as-a-Service agreement with Agility Robotics to use its bipedal robot, Digit. This, too, is a commercial contract that, following trials, is moving into manufacturing, parts supply, and logistics work. With Walden, Toyota spun off a research organization and injected capital into it; with Agility, Toyota is bringing in an external product as a service. The two cases show that Toyota is introducing different types of machines and different contract structures into its own factories.
Given that Toyota is pursuing multiple approaches, Walden, too, will need to demonstrate cost-effectiveness against existing equipment and third-party robots. Can it be integrated into existing processes quickly? Can learned skills be reused across different factories? Can it outperform conventional equipment on total cost, including downtime? Once Walden discloses customer names and operational data, and once it can be deployed quickly at plants beyond Toyota's own, it will move beyond being seen as a bet on prior research achievements and instead be evaluated as reproducible industrial infrastructure.