Automatic Gait Optimization with Gaussian Process Regression
Abstract: Gait optimization is a basic yet challenging problem for both quadrupedal and bipedal robots. Although techniques for automating the process exist, most involve local function optimization procedures that suffer from three key drawbacks. Local optimization techniques are naturally plagued by local optima, make no use of the expensive gait evaluations once a local step is taken, and do not explicitly model noise in gait evaluation. These drawbacks increase the need for a large number of gait evaluations, making optimization slow, data inefficient, and manually intensive. We present a Bayesian approach based on Gaussian process regression that addresses all three drawbacks. It uses a global search strategy based on a posterior model inferred from all of the individual noisy evaluations. We demonstrate the technique on a quadruped robot, using it to optimize two different criteria: speed and smoothness. We show in both cases our technique requires dramatically fewer gait evaluations than state-of-the-art local gradient approaches.
Introduction
Tuning a robot's walk is a notoriously tedious task. Legged robots operate in high-dimensional control spaces where optimal parameters vary based on surface conditions, hardware wear, and specific task requirements. While automated methods like evolutionary algorithms or local gradient descent (hill climbing) exist, they are often inefficient. These methods typically lack "memory" of past trials, get trapped in local optima, and struggle with the inherent noise of physical robot evaluations. Consequently, finding an optimal gait can require dozens of hours of physical trials. This paper introduces a Bayesian approach using Gaussian Process (GP) regression to find near-optimal walking parameters up to five times faster than traditional techniques.
Methodology
The researchers frame gait optimization as a sequential Bayesian optimization problem. They utilize a parameterized "walk engine" with 15 tunable variables, such as stride length and period. Instead of local perturbations, the relationship between these parameters and performance (velocity or smoothness) is modeled using a Gaussian Process with a Radial Basis Function (RBF) kernel.
The GP serves as a surrogate model that incorporates all past physical trials and explicitly models observation noise. To select the next set of parameters for testing, the system employs a "Most Probable Improvement" (MPI) acquisition function. This function searches the GP's posterior distribution to identify parameters most likely to exceed the current best performance, effectively shifting the computational burden from physical trials to internal modeling.
Experimental Setup
The algorithm was tested online using a Sony AIBO ERS-7 quadruped robot. The robot evaluated gaits by walking between visual landmarks in a physical arena. Velocity and smoothness (measured via head stability) were estimated through the robot's camera perception. This real-world setup ensures that the optimization accounts for the complexities and noise of physical hardware.
Key Results and Insights
The GP-based approach demonstrated significant improvements in data efficiency and performance over baselines. A key insight is that physical trials are expensive while memory is cheap. By conditioning every new guess on the entire history of evaluations, the GP avoids the "amnesia" of local methods.
- Velocity Optimization: The GP method achieved a competitive walk speed of 0.285 m/s in approximately 120 evaluations (roughly 2 robot-hours). This is a 5x speedup compared to state-of-the-art hill-climbing methods, which typically require 9 hours.
- Smoothness Optimization: For smoothness, the optimal gait was identified in just 20 observations (about 20 minutes), showcasing the method's rapid convergence.
- Robustness: As shown in the figure above, the GP approach consistently outperformed both random search and hill-climbing baselines, effectively bypassing local optima that hindered the gradient-based approach.
Conclusion
While the current model uses fixed hyperparameters, the results demonstrate that Bayesian optimization via Gaussian Processes is a powerful tool for robot gait tuning. Future work could involve kernel optimization, hyperparameter learning, and transfer learning—applying models across different robots or surfaces. This framework also holds potential for other continuous tuning problems, such as graph-cut stereo matching in computer vision.
Summary updated