Robots with more degrees of freedom than a task requires can reach the same goal in multiple ways, each with a distinct motion style. The challenge is to preserve a chosen style while also following an independent command. The CODES framework, introduced by the paper's authors, addresses this by first retargeting diverse recordings into motions a target robot can execute. It then uses weakly supervised consensus encoding and density-based clustering to discover a style repertoire without knowing in advance how many styles exist.

Once styles are discovered, reinforcement learning trains a policy that composes them with independent commands. The authors evaluated CODES on a Unitree Go2 quadruped in two simulators and on physical hardware. They report that CODES significantly improves kinematic coherence compared to baselines, and that the controller tracks velocity commands that go beyond the support of the recorded data while maintaining distinguishable style characteristics.

The work suggests that weakly supervised discovery can replace manual style specification, and the cross-simulator and real-world results indicate the learned styles transfer. The paper is available on arXiv.