Machine Learning for gait discovery
One of the central goals of the lab's research is improving the locomotive capabilities of the snake robots -- or more generally, improving the performance of teleoperated systems with complex, difficult to model dynamics. One approach is to use human intuition to modify the control of the robot, but this can be limited as the complexity of the system and environment grows. Furthermore, the snake robots have no analytic motion model that generalized well to all terrains, which prevents the straightforward application of many traditional robotic planning techniques and rigorous theoretic approaches such as geometric mechanics. Instead, we look to the successes in the field of machine learning to help tune our existing controllers and develop new ones. In doing so, we build off many of the core ideas and lessons learned from parallel research within our lab as well as others -- the virtual chassis for improved processing of feedback, insights about the need for gait-based approaches from geometric mechanics, and our extensions of the serpenoid-curve gait model from Hirose's pioneering work.
When optimizing the performance (speed, efficiency, robustness, etc.) of difficult-to-model systems such as the snake robot, sometimes the only method to test a set of control parameters is to evaluate the parameters on the real robot. Unfortunately, each of these evaluations can take considerable time, which can be limited when working with a physical robot. Therefore, this optimization must be performed with consideration for the limited number of available evaluations.
Furthermore, many traditional optimization methods (such as gradient ascent) assume that any objective function evaluation is paired with the value of the gradient at that point, whereas evaluations of physical systems or large simulations often do not provide this information.
We use principles from the field of Bayesian Optimization to sequentially optimize the locomotive performance of the snake robots. Rather than quickly choosing the next experiment parameters to test on the robot, all previous data is used to make a careful, statistical tradeoff between exploring unknown parameters and tuning well-performing parameters. Whereas typical non-expensive optimization approaches only take microseconds to choose parameters, but require many evaluations on the robot to find the best parameters; expensive optimization methods can take a few seconds to select these parameters but result in drastically fewer robot experiments to converge to the best parameters.
These methods consider all previous experimental data in the optimization by building a global approximation of the system's performance. This approximation includes both the expected value of the system performance as well as uncertainty in that estimate. For the subsequent tests on the robot, the paramter value which maximizes a statistical tradeoff such as Expected Improvement is chosen.
Improving Robot Performance
By applying these machine learning techniques to the robot, we are able to successfully optimize the performance of our existing gaits for various tasks, such as sidewinding locomotion up a steep slope. We have also used extensions of our work to tackle more complex challenges, such as optimizing an controller that can adapt to varying environmental conditions, and finding the Pareto optimal solutions for the multi-objective task of maximizing head stability while simultaneously maximizing locomotive speed.