Robots that can move around and manipulate objects—like picking up items or opening doors—are becoming increasingly important for industries ranging from healthcare to logistics. However, teaching these mobile manipulators to adapt and improve their skills in real-world settings remains a major challenge. This is because their control systems involve complex interactions between moving bases (the robot’s “body”) and arms (the “hands”), and fine-tuning these systems typically requires a lot of trial-and-error and time-consuming data collection. A newly published research paper introduces a novel approach called HiRe-MoMa that promises to make real-world robot learning faster, more stable, and more efficient, even with limited human guidance.
Key Takeaways
- HiRe-MoMa uses a hierarchical residual learning framework that fine-tunes existing robot control policies by adding small corrective adjustments, rather than relearning from scratch.
- The method separates corrections for the mobile base and the manipulator arm, reducing interference and better matching the robot’s physical structure.
- It incorporates sparse human interventions—less than 20% of training time—to guide the robot’s learning process more effectively.
- Real-world tests show significant improvements in robot task performance after just 30 minutes of online fine-tuning, regardless of the original control policy used.
At the heart of this research is the challenge of mobile manipulation (MoMa), which involves coordinating a robot’s movement and its ability to interact with objects. Traditional approaches often require training complex control policies from scratch, which can be slow and unstable when transferred to real robots. The HiRe-MoMa framework tackles this by building on a “base policy” that is already trained but may not perform optimally in a new or changing environment. Instead of replacing this base policy, HiRe-MoMa learns a “residual” policy—a set of small corrections that improve the base behavior.
To manage the complexity of the robot’s control system, the researchers designed a Hierarchical Residual Architecture. This means the residual corrections are split into two parts: one focused on the mobile base, and the other on the manipulator arm. These two parts interact in a one-way manner, which helps prevent the corrections from conflicting with each other during learning. This is important because the mobile base and arm have different physical roles and constraints, and aligning the learning structure with this hierarchy makes the fine-tuning process more stable and meaningful.
Another key innovation is the use of Expert-Guided Value Alignment. In simple terms, the robot occasionally receives sparse guidance from a human expert during training—less than 20% of the time. This guidance helps reshape the robot’s internal evaluation of actions (known as the “critic’s value landscape”), steering the robot towards more promising behaviors without needing constant supervision. This approach significantly improves sample efficiency, meaning the robot learns more from less data, which is crucial when training on physical machines where data collection is costly and time-consuming.
The team validated HiRe-MoMa on the TIAGo++ robot platform, a mobile manipulator used in research and industry. Their experiments showed consistent improvements in task success rates after only about 30 minutes of real-world interaction, demonstrating the framework’s practicality and efficiency. Notably, the method worked well regardless of the original base policy’s design, suggesting broad applicability across different robotic systems.
Looking ahead, this research opens the door to more adaptive and responsive robotic systems that can quickly fine-tune themselves on-site with minimal human input. Such capabilities are essential for deploying robots in dynamic, human-centered environments like hospitals, warehouses, or homes, where conditions and tasks often change. While challenges remain—such as scaling to even more complex tasks and ensuring safety during learning—the HiRe-MoMa framework offers a promising step toward more practical and intelligent mobile manipulation. For those interested in the technical details and ongoing developments, the researchers have made their project resources available online.
Based on research published on arXiv by Zhefei Gong, Hefeifei Jiang, Zehao Wang et al..