The key to teaching a robot five different gaits turned out to be knowing when to switch the AI trainer off. A team of researchers has shown that the same tool that makes walking natural actively prevents a robot from learning to run and jump – and that leaving it out for those two skills is what makes the whole system work.
For most of robotics history, teaching a robot to walk was one job. Teaching it to run was another. Teaching it to climb stairs was a third. Each required its own specialised controller, its own training process, its own carefully tuned rules. The idea that one brain could handle all of them had always been the goal — but getting there reliably, on a real robot, without feeding it thousands of hours of human motion capture data, had remained out of reach.
A paper posted to the arXiv research server on 22 April 2026 describes a system that does exactly that. A humanoid robot with twelve degrees of freedom has learned to walk, goose-step, run, climb stairs, and jump — five completely different ways of moving — using a single reinforcement learning policy with a consistent control structure and reward formulation. The same AI brain runs all five gaits. When the task changes, the brain switches modes rather than handing off to a different system entirely.
The key insight is knowing when to use a particular training tool and when to leave it out.
Reinforcement learning for robot movement often uses something called an Adversarial Motion Prior, or AMP. The idea is to show the AI a library of natural-looking movement examples and train a second network that acts as a critic, penalising the robot when its movements look unnatural.
This makes the resulting motions smoother and more stable. But the researchers found that AMP works well for some gaits and actively gets in the way for others.
For walking, goose-stepping, and stair climbing, the adversarial critic is helpful. These are periodic, stability-focused gaits where looking natural also means moving safely. The critic pushes the robot toward smooth, predictable patterns and that is exactly what you want.
For running and jumping, the same critic becomes a problem. These are explosive, dynamic movements where the robot needs freedom to find solutions that might look unusual but work mechanically. Applying the adversarial critic to these gaits over-constrains the motion, preventing the robot from discovering the movements that actually succeed. So the researchers left AMP switched off for those two gaits entirely.
The result is a robot that learned five gaits together, trained in simulation, and then transferred its behaviour directly to a physical robot in the real world without any additional real-world training — what roboticists call zero-shot sim-to-real transfer. The policy was validated on hardware across all five modes.
This matters because humanoid robots are increasingly being positioned as general-purpose physical workers. The companies building them are betting that a single robot platform can handle warehouses, construction sites, care homes, and disaster response. Each of those environments demands different kinds of movement. A warehouse requires smooth, efficient walking. Stairs appear in almost every building. Emergencies might require running. A system that handles all of these from one unified controller, without requiring a different specialist program for each terrain, is meaningfully closer to that goal.
The paper also eliminates the need for motion capture data entirely. Previous approaches to making robot movement look natural typically required large libraries of recorded human movement, which are expensive and time-consuming to collect. The reward functions in this system are designed from biomechanical principles instead — things like keeping the knee straight during the support phase of a stride, or swinging the arms in the opposite direction to the legs. The robot learns naturally looking movement by being rewarded for biomechanical correctness, not by being shown examples.
The paper has been submitted to the 2026 IEEE International Conference on Robotics and Automation. It is available as a preprint on arXiv and has not yet completed peer review.
Sources
Jiang, F., Chen, Y., Xu, K., Liu, Y., Wang, H., Shen, Z., Lu, J., Huang, S., Wang, Y., Xie, C. and Wu, R. “Multi-Gait Learning for Humanoid Robots Using Reinforcement Learning with Selective Adversarial Motion Prior.” arXiv cs.RO. Submitted 22 April 2026.
Peng, X.B., Ma, Z., Abbeel, P., Levine, S. and Kanazawa, A. “AMP: Adversarial Motion Priors for Stylized Physics-Based Character Control.” ACM Transactions on Graphics, vol. 40, no. 4, 2021. https://arxiv.org/abs/2104.02180
More Like This
Ray Jackson holds a BSc in Electrical Engineering from the University of Manitoba and a PhD in Physics from Carleton University. His reporting interests include Current and Future Technologies, Engineering and Artificial Intelligence.