UC Berkeley

Once robots are deployed in the real world, they will inevitably encounter scenarios that they have never seen before. Consequently, to develop robots that can help us in routine tasks, we need to first make progress towards the problem of generalization in robotics, which asks the question: how do we get robots to handle scenarios beyond what they have seen before? This thesis will present an approach to achieve such generalization through the use of rapid adaptation. Concretely, the proposed algorithm - Rapid Motor Adaptation (RMA) - allows robots to generalize by adapting to new and unseen scenarios in fractions of a second. The proposed algorithm is developed in the context of blind quadruped walking in complex terrain in the real world. Subsequently, it is extended to quadruped walking with egocentric vision, enabling robots to cross very challenging terrains, such as stepping stones, which are beyond the reach of blind systems. The same algorithm is then applied to the task of in-hand rotation to develop a single controller capable of rotating a diverse set of objects in the real world, as well as to bipedal robots.