A new preprint introduces SOMA, a hierarchical reinforcement learning framework designed to let quadruped robots manipulate stacks of objects without needing a dedicated robotic arm. The work targets heavy-duty physical manipulation in unpredictable environments, a setting where legged robots often struggle because they lack the specialized hardware used for grasping or pushing.
The framework is presented as a way to coordinate whole-body actions, allowing the robot to use its own body and legs to interact with stacked objects. Because the paper is a single arXiv preprint, the claims are not yet peer-reviewed or compared with other methods in the summary provided.
SOMA's significance lies in extending quadruped capabilities beyond locomotion into physical manipulation, potentially enabling tasks like clearing debris or rearranging materials in unstructured settings. The authors position this as a step toward more versatile legged robots that can operate where traditional manipulators are impractical.