Toyota spins off a robotics unicorn: How does Walden bring a “big behavior model” into factories?

Author: Zen, PANews

Before July 15, Walden Robotics had been largely unknown to the public.

And on that day, the robotics company—spun out of the Toyota Research Institute—suddenly made its official debut, disclosing in one go $300 million in seed round funding and an $1.1 billion valuation.

The round was co-led by Toyota and Deviation Capital, with industry investors including NVIDIA, Boeing, Samsung Ventures, Prologis Ventures, and CoreWeave Ventures participating.

From its founding to joining the unicorn ranks, Walden only took half a year. And it already has many of the conditions that robot startups dream of—an established research team, ample capital, Toyota’s open manufacturing system, and potential partnership channels from investors in manufacturing, aviation, electronics, and logistics.

The new unicorn that emerged from Toyota Research Institute

Before the funding news was released, Walden Robotics had been in stealth mode.

In January this year, Walden was spun out of Toyota Research Institute (hereafter “TRI”) to establish an independent company. The company name is inspired by American writer Henry David Thoreau’s book Walden, which discusses the importance of conscious, purposeful living. It also corresponds to the questions the company hopes to explore: how robots can help people find more meaning in work and life.

According to Walden co-founder and CEO Russ Tedrake, a general-purpose robot powered by physical AI is undoubtedly a disruptive technology that has already reached a critical turning point. However, for commercial success, robot companies still need to validate unit economics and collaborate deeply with customers.

After becoming an independent company, Walden can focus more on commercializing robotics technology from Toyota Research Institute, taking relevant results from the lab into production environments. By partnering with large global manufacturing and logistics enterprises, Walden hopes to continuously validate product capabilities in real-world scenarios, ensure the product fits actual production processes, and deliver clear cost savings and efficiency improvements.

Russ Tedrake is a professor at MIT. Previously, he led the robotics and machine learning teams at TRI for nearly a decade. His team has made many contributions to foundational research, including Diffusion Policy, Universal Manipulation Interface (UMI), Large Behavior Models, OpenVLA, and the open-source simulator Drake.

In addition to Russ Tedrake, Walden’s founding team currently also includes CTO Ben Burchfiel, COO Kerri Fetzer-Borelli, Chief Product Officer Dave Johnson, Chief Strategy Officer Adrien Gaidon, Chief Architect Siyuan Feng, and AI lead Rares Ambrus. Many of them are also project leads in TRI’s Large Behavior Models research, participating in building model architecture, training, simulation, and evaluation systems.

Walden Robotics team; Russ Tedrake is the second from the left

It’s clear that, compared with ordinary startups, Walden’s starting point and platform are significantly higher. On one hand, it builds on decades of TRI research achievements in robotics. On the other hand, Toyota is not only its core investor, but also its most important early industrial cooperation partner, providing the first real production deployment scenarios.

Backed by Toyota’s manufacturing system, Walden shortens the commercial validation cycle

A common challenge for embodied intelligence companies is a gap between technology R&D and commercial deployment.

Robots need to enter real environments to obtain high-quality data, but early products suffer from issues with reliability and economics that make it difficult for enterprise customers to be convinced and deploy them in actual work. Without deployment scenarios and data, models struggle to cover real-world anomalies, and product capabilities also become hard to improve continuously.

But Walden, which received support from Toyota’s production system from the outset of its founding, has shortened this validation cycle to some extent. Toyota is both its technology incubator and core investor, as well as the provider of the first real deployment scenarios. Walden does not need to start from scratch to find industrial customers, nor does it need to separately build a simulated factory for testing; it can instead enter existing production workflows, define tasks together with the manufacturing team, adjust equipment, and evaluate input-output performance.

The value of this industrial background goes beyond simply providing a “training”场地 for robots. Whether industrial robots can create economic value depends on multiple factors such as task frequency, equipment utilization rate, and safety requirements. Robot tasks that perform well in laboratories may not necessarily have deployment value once they enter factories.

Meanwhile, Toyota’s long-accumulated experience in manufacturing and automation can help Walden prioritize the operations best suited to its technical capabilities at the current stage—while also having clear commercial returns—reducing the risk that product development becomes misaligned with customer needs.

In addition, Walden’s investor roster also provides potential channels for expanding external scenarios. Beyond Toyota, Boeing, Samsung Ventures, and Prologis Ventures correspond to aviation manufacturing, the electronics industry, and logistics infrastructure, respectively. NVIDIA and CoreWeave connect robot computing and AI training resources.

These companies are clearly potential synergy resources, and are expected to offer collaboration entry points to Walden in the future. To a certain extent, after Toyota solves Walden’s issues with scenarios and data in the very early commercial stage, what truly determines Walden’s long-term value may be whether this technology and operations system can move beyond Toyota and be converted into standardized products for more manufacturing enterprises.

On this point, Walden is highly confident, inheriting TRI’s research and technological achievements—so it’s necessary to mention the core of the company’s technology stack: Large Behavior Models (LBM).

Core technology LBM (Large Behavior Models) brings general-purpose manipulation capability into factories

Unlike large language models aimed at text generation, LBM needs to simultaneously process visual inputs, the robot’s own state, tactile or other sensor information, and task instructions, and then generate continuous actions accordingly. Its goal is not to write a separate program for each job; instead, through multi-task data training, the same model learns and transfers different manipulation skills.

This roadmap is built on years of TRI research in robot learning. Among the representative technologies is Diffusion Policy.

Conventional industrial robots typically rely on pre-set motion trajectories and workcell conditions. When part positions, equipment layouts, or production processes change, engineers often have to reprogram and debug again. Diffusion Policy learns the distribution of actions from human demonstrations; the model extracts patterns from visual, action, and robot state data, and then attempts to autonomously reproduce them.

On top of that, LBM further incorporates multiple tasks into a unified pretraining framework. TRI’s previously disclosed research used close to 1,700 hours of robot data and conducted 1,800 tests in real environments and more than 47k simulation tests. The research results show that, when learning parts of new tasks, models pretrained on multiple tasks require significantly less data than single-task models trained from scratch.

In both simulation and the real world, Walden evaluates its LBM models across different tasks and environmental conditions

This provides the foundation for Walden’s product logic: robots don’t have to rely on engineering teams to program each step item by item; instead, they can adapt to new operation workflows through a small amount of demonstrations. For industrial customers, this capability is mainly applicable to manufacturing environments where there are many product varieties and production tasks need frequent adjustments. Compared with traditional automation equipment that can only repeat fixed motions, robots with learning ability are expected to switch processes and tasks with lower retrofit costs.

At present, Walden uses a combination of autonomous operation and remote human assistance. Robots can independently complete regular tasks they have already mastered; when encountering anomalies, environmental changes, or situations beyond the model’s capability range, remote operators step in.

In the robot’s body design, Walden adopts a humanoid upper body with a two-arm form combined with a wheeled mobile base. The product focus is placed on dual-arm manipulation, task learning, and environmental adaptation.

Wheeled mobile robots are not uncommon in industrial and warehousing scenarios where the ground is flat and workstations are well-defined. Their main advantages are stability, load capacity, and relatively controllable system complexity. The humanoid torso design helps the robot use tools designed for humans and operate within workspaces intended for humans. Its pursuit of “generality” comes more from the model’s ability to learn across different tasks, as well as the dual-arm system’s ability to handle and manipulate a variety of objects and equipment.

However, although Walden has unique advantages and is ahead to some extent in the robotics track, as Russ Tedrake said during Walden’s formal debut: “The team is strong enough and the progress is fast enough, so we don’t need to hype it.” But for a company that has just emerged from stealth mode, as Russ Tedrake also said: “We’re just starting this journey.”