Atlas is best known for highly visible displays of movement, including parkour and dance routines. Its latest advance is quieter, but potentially more important: the humanoid robot has learned to walk and grab objects using one artificial intelligence model.
That matters because learning robots have typically split these abilities across separate systems. One model might handle walking or jumping, while another manages grasping. The new work from Boston Dynamics and the Toyota Research Institute (TRI) brings arms and legs under a single learned controller.
Why One Model Matters
The model developed for Atlas is described as a generalist system. It learns from a range of example actions and then controls both the robot's arms and its legs. Russ Tedrake, a roboticist at the Toyota Research Institute and the Massachusetts Institute of Technology, led the current work, with Scott Kuindersma, VP of robotics research at Boston Dynamics, as co-lead.
Tedrake framed the shift in a simple way: “The feet are just like additional hands, in some sense, to the model,” he says. “And it works, which is just awesome.”
The point is not only that Atlas can perform multiple movements. The more significant claim is that one model can coordinate them in a way that appears more unified. For a humanoid robot, that coordination is central. Reaching for an object is not only an arm motion; the body has to stay balanced, the legs may need to shift, and the robot has to keep track of its own position as it moves.
What The Atlas Model Takes In
The single model receives several kinds of input. It uses images from the robot's visual sensors, proprioception data from bodily sensors, and language prompts linked to different actions. Proprioception gives the robot a continuous sense of its own position and movement.
The model is trained through examples of Atlas performing different tasks. Those examples come from a mix of teleoperation, simulation, and demonstration videos. The result is a large behavior model (LBM) that controls the humanoid robot in a way that can look more natural than a set of isolated task systems.
One example is bin picking. When Atlas reaches low into a bin, the robot may reposition its legs to rebalance, much as a person would. That kind of movement is important because it joins manipulation and locomotion into the same action instead of treating them as separate problems.
The Question Of Emergent Behavior
The LBM has also shown what the source describes as basic emergent behavior. When Atlas drops an item, it can bend down and pick it up, displaying a recovery skill it was not specifically trained to perform.
This is the part that makes the work feel connected to the broader arc of artificial intelligence. Large language models (LLMs), trained on large amounts of text data, have sometimes shown unexpected abilities, including the ability to code. Roboticists hope a similar approach could lead to machines that discover useful new skills while trying to complete real-world tasks.
Tedrake says Atlas and other robots are beginning to show signs of more generalized learning. His lab is also testing different kinds of robot arms trained on tasks such as slicing vegetables and sweeping up spilled coffee beans.
He argues that the evidence so far points in the same direction as the approaches used for LLMs. “I think it's changing everything,” he says.
Why Robot Videos Need Careful Reading
Progress in robotics can be hard to judge from short clips alone. Videos of commercial humanoids may show robots loading refrigerators or taking out the trash with apparent ease. But the source notes that such clips can be misleading.
Humanoid robots are often teleoperated, programmed in advance, or trained for one task in controlled conditions. That does not make the demonstrations meaningless, but it does mean viewers should ask what is being shown, what was trained, and how reliably the system works outside the clip.
Ken Goldberg, a roboticist at UC Berkeley who receives some funding from TRI but was not involved with the Atlas work, calls the new work meaningful. “It's definitely a step forward,” he says. “The coordination of legs and arms is a big deal.”
Goldberg also urges caution around the word emergent. As with large language models, surprising behavior can sometimes be traced back to examples that appeared in training data. In robotics, a skill may look more novel than it really is unless researchers explain how often the robot succeeds and how it fails during experiments.
TRI has previously been transparent with its work on LBMs, and the source says it may release more data on the new model. That kind of detail would help separate impressive demonstrations from durable progress.
What Comes Next For General Robot Learning
The larger open question is whether simply increasing the data used to train robot models will unlock more emergent behavior. At a debate held in May at the International Conference on Robotics and Automation in Atlanta, Goldberg and others cautioned that engineering methods will also remain important.
Still, the Atlas work suggests a possible path toward robots that learn in a broader way. If models can connect vision, body awareness, language prompts, and examples of action, future robots may need less narrow retraining for each new task.
The source points to a long-term possibility: robots that can work in messy environments and quickly learn new skills, from welding pipes to making espressos. That is not presented as something already solved. It is the direction researchers are watching as humanoid robots and other machines move closer to real-world use.
Tedrake believes robotics is approaching a turning point. “I think we need to put these robots out of the world and start doing real work,” he says.