Learning Whole-Body Manipulation for Quadrupedal Robot

We propose a learning-based system for enablingquadrupedal robots to manipulate large, heavy objects using theiwhole body. Our system is based on a hierarchical control strategy that uses the deep latent variable embedding which capturesmanipulation-relevant information from interactions, propriocep-tion, and action history, allowing the robot to implicitly understandobject properties. We evaluate our framework in both simulationand real-world scenarios. In the simulation, it achieves a successrate of 93.6%in accurately re-positioning and re-orienting variousobjects within a tolerance of 0.03 m and 5 degrees. Real-world experimentsdemonstratethesuccessfulmanipulationofobjectssuchasa19.2kgwater-filled drum and a 15.3 kg plastic box filled with heavy objects while the robot weighs 27 kg. Unlike previous works that focus onmani pulating small and light objects using prehensile manipulation, our framework illustrates the possibility of using quadrupeds for manipulating large and heavy objects that are ungraspable with the robot's entire body. Our method does not require explicit object modeling and offers significant computational efficiency comparedto optimization-based methods.