Teaching robots to exploit multiple symmetries at once

Robots naturally exhibit multiple symmetries from their mechanical structure and task design, yet current approaches treat each separately and miss their combined potential. This work introduces a framework that makes robot policies jointly equivariant to multiple symmetries by mapping them into a unified representation space, mixing symmetries from both the robot's configuration and the task itself. Experiments on a dual-arm robot show the approach improves generalization on unseen tasks and configurations.