MobiAgent: Dual-Loop Recursive Policy Self-Improvement for Long-Horizon Mobile Manipulation
This preprint introduces MobiAgent, a dual-loop framework for long-horizon mobile manipulation. It decouples high-level VLM planning from low-level flow-matching skills and adds an automated lifelong learning loop from deployment data. The abstract reports no quantitative results, but the approach matters because it targets compounding errors and interference between locomotion and arm control.
How can long-horizon mobile manipulation agents combine robust hierarchical execution with continuous, automated self-improvement?
Long-horizon mobile manipulation suffers from compounding execution errors and capacity interference between locomotion and arm control. Existing Vision-Language-Action models handle short-horizon tasks but lack hierarchical reasoning for multi-stage objectives, while previous hierarchical agents use rigid sub-task mapping, inflexible replanning, and no continuous learning.
Prior methods include Vision-Language-Action models and hierarchical agent architectures. VLAs perform well on short-horizon tasks but do not provide the hierarchical reasoning needed for multi-stage objectives; hierarchical agents suffer from rigid sub-task mapping, inflexible replanning, and lack of continuous learning.
MobiAgent is a dual-loop agentic framework. The Inner Loop separates high-level reasoning from low-level control using composable atomic skills, with a Vision-Language Model for receding-horizon planning and visual reflection, dynamically composing skills for error recovery. Skills are executed by specialized flow-matching experts that share a unified VLM backbone to maximize reusability and reduce capacity interference. The Outer Loop drives automated lifelong learning by autonomously segmenting, verifying, and clustering deployment rollouts for recursive policy improvement.
Not reported in the abstract.
Not reported in the abstract. No quantitative evaluation, robot platforms, or code release details are provided.
Warehouse automation, logistics, and service-robot makers could potentially use this approach for long-horizon mobile manipulation. Adoption is likely 1–3 years away if real-robot validation, safety guarantees for the continuous learning loop, and evidence of reliability over long deployments are established first.
The full text is not republished here because the paper's license does not allow it. Read the original on arXiv.