Boston Dynamics and Toyota Showcase Atlas Robot’s New AI Skills

Boston Dynamics and the Toyota Research Institute (TRI) have unveiled a significant advancement in robotics with the demonstration of their Atlas humanoid robot, powered by a new Large Behavior Model (LBM). This breakthrough, showcased in a video released on August 26, 2025, highlights the robot’s ability to perform a series of complex tasks that integrate both object manipulation and locomotion.

Atlas is now capable of executing a continuous sequence of actions, including walking, crouching, and lifting, to complete tasks such as packing, sorting, and organizing. The introduction of LBMs allows for rapid development of new capabilities without the need for extensive programming, enabling the robot to adapt quickly to unexpected challenges. For instance, during the demonstration, Atlas successfully navigated physical obstacles, adjusting its movements in real-time to close a box lid and slide it across the floor.

Innovative Control Through Large Behavior Models

Traditionally, humanoid robots have managed walking and manipulation tasks separately, relying on distinct control systems for locomotion and arm movements. In contrast, the new approach utilizes a single LBM that governs the entire robot, treating the hands and feet as part of a unified system. This innovation stems from a collaborative research partnership established in October 2024 between Boston Dynamics and TRI, aimed at harnessing their combined expertise to enhance the capabilities of smart robots.

Scott Kuindersma, Vice President of Robotics Research at Boston Dynamics, emphasized the significance of this work, stating, “This work provides a glimpse into how we’re thinking about building general-purpose robots that will transform how we live and work.” He noted that training a single neural network on various long-duration tasks can lead to improved generalization, positioning Atlas as a leading example of a highly capable robot.

Russ Tedrake, Senior Vice President of Large Behavior Models at Toyota Research Institute, added that the true potential of humanoid robots lies in their ability to perform diverse tasks in real-world environments. He remarked, “Large Behavior Models address this opportunity in a fundamentally new way – skills are added quickly via demonstrations from humans, and as the LBMs get stronger, they require less and less demonstrations to achieve more and more robust behaviors.”

Understanding Large Behavior Models

The term LBM refers to an emerging category of artificial intelligence designed to learn and execute physical actions and decision-making processes in real-world scenarios. Unlike Large Language Models (LLMs), which are trained on extensive text datasets to understand and generate human-like language, LBMs focus on predicting and executing behaviors. Examples of successful LBMs include TRI’s manipulation systems and Nvidia’s Isaac GR00T, which are tailored for embodied AI applications.

As research continues, the collaboration between Boston Dynamics and TRI aims to address foundational questions regarding humanoid robots and LBMs. The insights gained from this project could pave the way for further advancements in the field, enhancing our understanding of large models for whole-body control and advanced manipulation techniques.

This innovative demonstration serves as a testament to the transformative potential of artificial intelligence in robotics, offering a glimpse into a future where humanoid robots may play an integral role in our daily lives.