Transfer Learning for Faster Tensegrity Robot Gait Optimization

Tensegrity Robots are a class of soft robots gaining attention due to their ease-of-assembly, ability to preserve structural integrity under physical deformation, and ability to move with dynamic gaits. Previous work has shown how optimization techniques can be used to find the optimal gait in a given environment to move as efficiently as possible. The problem with these techniques, however, is that they suffer from the "Cold Start," where they need to re-learn the optimal gait for the robot when it is deployed in new, unseen conditions and environments. Our research explores how we can use Transfer Learning to leverage knowledge from previously seen and learnt environment, called Source Task, to a new environment, called Target Task, in order to learn these gaits faster. We use Bayesian Optimization as our base optimization technique and a Transfer Learning Framework that models the distance between the Source and Target tasks. This approach shows an improvement in both the performance of the learned gait in the Target Task and the number of optimization trials it takes to learn this optimal gait.