Robotics
New AI approach helps legged robots navigate rough terrain without rigid gait constraints
Researchers propose a reinforcement learning method that guides robot leg movements through reward signals rather than preset motion templates, improving adaptability across different robot platforms.
A team of researchers has introduced Phase-Guided Terrain Traversal (PGTT), a machine learning technique designed to help legged robots traverse uneven ground more reliably by using a flexible approach to controlling leg movement patterns.
The Problem with Current Methods#
Existing perceptual AI controllers for legged robots face a difficult tradeoff. Some systems rely on oscillator-based motion blueprints or inverse-kinematics priors—essentially hardcoded movement patterns that limit how the robot can adapt. Others operate with minimal terrain awareness, making them vulnerable to unexpected ground conditions and sensor noise while unable to anticipate what's coming under the rear legs.
How PGTT Works#
Rather than imposing rigid movement templates, PGTT uses deep reinforcement learning with reward-shaped guidance to encourage structured gaits. The system encodes each leg's motion phase as a cubic Hermite spline, automatically adjusts how high the leg swings based on local terrain features, and penalizes unwanted contact during swing phases. Critically, the robot's policy operates directly in joint-space commands, making it agnostic to different robot body designs.
Real-World Performance#
The researchers trained PGTT in physics simulation (MuJoCo) on procedurally varied stair-like terrain with curriculum learning and domain randomization. In testing, PGTT achieved the highest success rate under push disturbances (+7.5% better than the next-best competitor) and on discrete obstacles (+9%) while maintaining similar velocity tracking performance.
The team validated the approach on a Unitree Go2 quadruped using a real-time LiDAR sensor pipeline and reported preliminary results on an ANYmal-C using identical hyperparameters—early evidence that the method can generalize across platforms without platform-specific tuning.
Significance#
The work suggests that steering robot behavior through thoughtfully designed rewards, rather than constraining action spaces upfront, can preserve learning flexibility while maintaining structured movement patterns. This may enable more adaptable locomotion systems across different robot morphologies.
