Boston Dynamics helped define what legged robots could look like in public imagination: machines that could run, dance, stack shelves, and perform parkour with a level of physical control that once seemed out of reach. Now Marc Raibert, the founder of Boston Dynamics, is pushing toward a different milestone: robots that can learn more of their own behavior instead of relying on every motion being engineered by people.
The shift matters because the field around Boston Dynamics has become crowded. Robot dogs and humanoids are no longer rare demos. The harder question is whether these machines can move beyond staged capability and become more independent in the real world.
Why robot learning is becoming the real contest
Boston Dynamics is no longer alone in showing machines with legs. A startup called Figure showed a humanoid called Helix that can apparently unload groceries. Another company, x1, showed a muscly-looking humanoid called NEO Gamma doing chores around the home. Apptronik said it plans to scale up manufacturing of its humanoid, Apollo.
Those demonstrations point to a market full of ambition, but they also raise an important caution. Demos can be misleading. Few companies disclose how much their humanoids cost, and it remains unclear how many truly expect to sell them as home helpers.
That is why the central test is not just whether a robot can complete an impressive routine once. The more important measure is how much it can do without direct human programming or remote control. A machine that needs a carefully scripted path for each new action is very different from one that can learn useful behavior through experience.
Raibert described the goal plainly:
“The hope is that we'll be able to produce lots of behavior without having to handcraft everything that robots do,”
How reinforcement learning changes the training process
The method behind this push is reinforcement learning, an artificial intelligence technique that lets a computer learn through experimentation combined with positive or negative feedback. It has been around for decades, but it gained wider attention last decade when Google DeepMind showed that it could produce algorithms capable of superhuman strategy and gameplay.
More recently, AI engineers have also used reinforcement learning to shape the behavior of large language models. For robots, the technique has a different purpose: helping physical machines discover how to move more effectively.
Raibert says the RAI Institute used reinforcement learning to improve Spot, Boston Dynamics' four-legged robot. Spot is used on oil rigs, construction sites, and other places where wheels struggle with the terrain. The upgrade helped Spot run three times faster.
The same approach is also helping Atlas, Boston Dynamics' humanoid robot, walk more confidently, according to Raibert. That matters because walking on legs is not only about moving forward. It requires balance, correction, timing, and the ability to handle changing conditions.
Simulation makes practice less costly
One reason this work is becoming more practical is the quality of simulation. Raibert says highly accurate new simulations have made the learning process faster by allowing robots to practice moves in silico before those behaviors are tested on the physical machine.
That changes the economics and risk of robot training. Instead of needing every experiment to happen on expensive hardware, much of the trial and error can happen in software first. Raibert put the advantage this way:
“You don't have to get as much physical behavior from the robot [to generate] good performance,”
Al Rizzi, chief technology officer at the RAI Institute, framed the benefit in more practical terms. AI may not immediately give people robots that can do the dishes, but it can reduce the damage that happens when new behaviors are transferred to real machines. As Rizzi said:
“You break fewer robots when you actually come to run the thing on the physical machine,”
That is a critical step for legged robots. A wheeled machine can often be tested with fewer dramatic failures. A humanoid or quadruped, by contrast, has to manage falls, unstable surfaces, and constant motion. Better simulation gives researchers more room to improve behavior before hardware takes the risk.
Research momentum is spreading beyond Boston Dynamics
Boston Dynamics has spent decades building legged robots, drawing on Raibert's work on how animals balance dynamically through low-level control from the nervous system. That foundation helped produce machines known for nimble movement. But even with that history, more advanced behaviors have typically required careful programming or some type of human remote control.
The newer work aims to reduce that dependence. In 2022, Raibert founded the Robotics and AI (RAI) Institute to explore ways of increasing the intelligence of legged and other robots so they can do more on their own.
Academic researchers are also showing how reinforcement learning can improve legged locomotion. A team at UC Berkeley used the approach to train a humanoid to walk around their campus. Another group at ETH Zurich is using the method to guide quadrupeds across treacherous ground.
Taken together, these examples suggest that the next phase of robotics may be less about making machines look impressive in isolated clips and more about teaching them how to adapt. Boston Dynamics helped make legged robots famous for physical skill. The larger challenge now is whether machine learning can make those robots more capable, more flexible, and less dependent on people scripting each step.