This Robot Only Needs a Single AI Model to Master Humanlike Movements (www.wired.com)

Boston Dynamics and the Toyota Research Institute have developed a groundbreaking single AI model that enables their humanoid robot Atlas to both walk and manipulate objects seamlessly—a notable departure from the traditional approach of using separate models for locomotion and grasping. This unified model, called the large behavior model (LBM), integrates visual inputs, proprioceptive data, and language prompts to generate coordinated, humanlike movements. Impressively, Atlas exhibits emergent behaviors such as instinctively recovering dropped items without explicit training, hinting at the potential for robots to autonomously adapt in complex environments. This advancement is significant for the AI and robotics fields because it mirrors the trajectory seen in large language models (LLMs), where scaling data and training methods have unlocked unexpected, generalized capabilities. By training Atlas on diverse examples from teleoperation, simulation, and demonstration videos, researchers have created a more versatile and natural form of robotic control. The work suggests that robots could soon handle a variety of unstructured tasks—ranging from manual trades to delicate household chores—without requiring exhaustive retraining for each activity. While experts caution that emergent robotic behaviors need careful validation to understand their novelty and reliability, this unified model approach marks a pivotal step toward more adaptive, versatile robots. As robotics faces an inflection point akin to the breakthroughs in generative AI, this development moves us closer to humanoid robots that can perform real-world tasks fluidly, opening exciting new possibilities for automated assistance in everyday and industrial settings.