Mobile manipulation (MoMa) is a key capability for autonomous robots, but fine-tuning such systems in the real world remains difficult. The high-dimensional action spaces and the intricate coupling between heterogeneous modules make direct learning impractical, according to a new paper on arXiv.

The paper introduces hierarchical residual policy learning as a way to tackle this problem. Instead of learning a single monolithic policy, the method builds policies at multiple levels, with residual corrections applied at each layer. Sparse human guidance is used to provide targeted feedback, reducing the need for dense supervision.

The authors position this as a step toward more practical real-world deployment of mobile manipulators. By structuring learning hierarchically and relying on occasional human input, the approach aims to lower the cost and complexity of fine-tuning in unstructured environments.