Boston Dynamics’ Atlas humanoid robot has achieved a major breakthrough: it now uses a single artificial intelligence model to control both walking and object manipulation — a significant departure from traditional robotics that requires separate models for each task.
One Model to Rule Them All
Developed in partnership with the Toyota Research Institute (TRI), Atlas’s new whole-body learning approach treats legs and arms as part of a unified system. Instead of relying on one AI model for locomotion and another for grasping, the robot learns both from a shared dataset of example behaviors.
“The feet are just like extra hands, in some sense, to the model,” explains Russ Tedrake, a roboticist at TRI and MIT who led the research. “And it works, which is just amazing.”
How the Learning Model Works
The single model receives input from the robot’s visual sensors, proprioception data (giving it continuous awareness of its position and motion), and language prompts linked to various actions. The model is trained on examples of Atlas performing tasks through a combination of teleoperation, simulation, and demonstration videos.
The resulting Locomotion and Behavior Model (LBM) controls the humanoid robot in a more natural-seeming manner. When picking objects out of a bin, for instance, the robot instinctively repositions its legs — much like a human would — to rebalance when reaching down low.
Emergent Behaviors
Perhaps the most exciting aspect is the emergence of untaught skills. When Atlas drops an item, it demonstrates a spontaneous “recovery” ability by bending down to pick it up again — a behavior it was never explicitly trained to perform.
This mirrors the way large language models (LLMs) sometimes exhibit unexpected capabilities, such as learning to code, when trained on massive datasets. Roboticists hope that similar scaling approaches will lead to robots that spontaneously develop diverse new skills when attempting to accomplish tasks.
Industry Implications
While polished demo videos showing humanoids performing complex chores are increasingly common, many of these rely on teleoperation or careful pre-programming. The Atlas work represents a genuine step toward robots that can operate in messy, real-world environments and learn new skills — from welding pipes to making coffee — without extensive retraining.
“It is definitely a step ahead,” says Ken Goldberg, a roboticist at UC Berkeley. “The coordination of legs and arms is a big deal.”
Goldberg cautions, however, that claims of emergent robotic behavior must be examined carefully. Just as some capabilities of LLMs can sometimes be traced to examples in their training data, robots may demonstrate skills that appear more novel than they actually are.
What’s Next
Whether simply scaling up the data used to train robotic models will unlock ever-more emergent behavior remains an open question. Tedrake’s lab is also experimenting with different types of robotic hands trained to perform diverse tasks, including cutting vegetables and sweeping up spilled coffee beans.
“I think it is changing everything,” Tedrake says of the unified model approach applied to robotics.
For an industry often caught between impressive demo videos and real-world limitations, the Atlas LBM represents meaningful progress toward truly adaptable, general-purpose humanoid robots.
