HiRe-MoMa may improve robot tasks after around 30 minutes online

Researchers Zhefei Gong, Hefeifei Jiang, Zehao Wang and Jiachen Li submitted an arXiv preprint on Oct. 14 describing HiRe-MoMa, which they said improved TIAGo++ robot tasks after around 30 minutes of online interaction.

The work matters because mobile manipulation combines navigation and arm control, making real-world adaptation costly; the paper said sparse human guidance covered under 20% of training rather than continuous supervision.

HiRe-MoMa keeps a base robot policy frozen and learns residual corrections, meaning small add-on adjustments to an existing controller. The authors split those corrections between the mobile base and manipulator arm, with one-way conditioning intended to match the robot's physical hierarchy.

The authors said the framework is policy-agnostic, meaning it can sit atop different base-policy architectures. In the abstract, they wrote that real-world experiments on TIAGo++ showed HiRe-MoMa "consistently enhances task performance with only around 30 minutes of online interaction."

The arXiv paper also describes Expert-Guided Value Alignment, a training method using sparse human interventions to steer autonomous exploration. The authors said those interventions used "under 20% training coverage" to reshape the value estimates that guide learning.

The evidence remains limited to what the authors reported in a preprint, which is not yet peer reviewed. The arXiv page provides no benchmark table, task list, failure rates, comparison metrics or deployment data in the supplied abstract, so the size of the improvement cannot be independently assessed from this material.

The arXiv listing names the subject areas as Robotics, Artificial Intelligence and Machine Learning and lists the paper as associated with CoRL 2026. It also links to a project website for further details.

Source: arXiv preprint, arXiv, Oct. 14, 2026