Boston Dynamics' Robot Dog Just Learned to Do a Handstand

Training robots today mostly happens by trial and error. Engineers let a simulated robot try a movement thousands of times, and each attempt teaches it a little more — much like a toddler learning to walk by falling down repeatedly. But some skills are so unforgiving that the robot fails in the first split second, every single time. It never gets far enough to learn anything. Researchers call these "narrow-viability" tasks, and they have been a stubborn wall.

The team's solution is elegantly simple. Instead of letting the robot's joints behave exactly as they do in real life, they start with the joints very stiff — like a person wearing a rigid back brace, or a bike with training wheels. A stiff robot doesn't topple as easily, so it survives longer during practice and actually collects useful feedback. As the robot's successful attempts get longer, the researchers slowly loosen the joints back to their true, real-world stiffness. By the end, the robot is learning under realistic conditions but with all the momentum of earlier easy wins.

They proved the idea on a simple balancing toy (a pole on a cart) and then on a genuine challenge: getting a four-legged Boston Dynamics Spot robot to rear up onto its front legs and hold a handstand. Trained the normal way, Spot always stalled at a policy that never completes the move. With the new curriculum, it succeeded across ten separate training runs in simulation — and the learned skill carried over to the physical robot.

The bigger takeaway is a new dial engineers can turn. When a robot is failing because it keeps getting knocked out early, not because the goal is unclear, adjusting how stiff or springy its simulated joints feel may be the fastest route to progress. Practically, that means fewer weeks of expensive trial and error before a robot can do something genuinely hard — and a shorter path from lab demo to useful helper.