Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This paper introduces a framework for learning robot policies equivariant to multiple symmetries across configuration and task spaces by composing symmetries that exist in different spaces. It leverages the differential-geometric structure of the forward kinematics map to descend symmetries from configuration to task space and lift symmetries from task to configuration space. Experiments on a dual-arm robot in simulation and the real world demonstrate that jointly leveraging multiple symmetries yields improved generalization compared to treating them in isolation.
Robots can learn faster and generalize better by composing symmetries across their configuration and task spaces, instead of treating them in isolation.
Robots exhibit a rich variety of symmetries arising from their mechanical structure and the properties of their tasks. Although many robotics problems exhibit several symmetries simultaneously, existing approaches typically treat them in isolation, failing to exploit their combined potential. This paper introduces cross-space symmetry compositions, a framework for learning robot policies that are jointly equivariant to multiple symmetries across configuration and task spaces. Leveraging the differential-geometric structure of the forward kinematics map, we both descend symmetries from configuration to task space and lift symmetries from task to configuration space, enabling their composition within a unified representation space. We validate our framework on simulated and real-world experiments on a dual-arm robot, demonstrating that jointly leveraging multiple symmetries yields improved generalization.