1 The University of Hong Kong2 Astribot3 Southern University of Science and Technology
* Equal contribution † Project leader ‡ Corresponding authors
It plans the task, executes it skill by skill,
and improves from its own deployment data.
ABSTRACT
Long-horizon mobile manipulation presents significant challenges due to compounding execution errors and capacity interference between locomotion and arm control. While recent Vision-Language-Action models excel at short-horizon tasks, they lack the hierarchical reasoning required for multi-stage objectives. Furthermore, existing hierarchical agents suffer from rigid sub-task mapping, inflexible replanning, and a lack of continuous learning. To address these limitations, we introduce MobiAgent, a dual-loop agentic framework that bridges robust deployment execution and recursive policy self-improvement. During deployment, the Inner Loop decouples high-level reasoning from low-level control through highly composable atomic skills. It employs Vision-Language models for receding-horizon planning and visual reflection, dynamically composing skills to ensure robust error recovery. These skills are executed by specialized flow-matching experts that share a unified VLM backbone, maximizing reusability while mitigating capacity interference. Concurrently, the Outer Loop drives automated lifelong learning by autonomously segmenting and verifying deployment rollouts, clustering them to discover atomic skills, and continuously fine-tuning the skill library without human annotations. Evaluations on RoboCasa, BEHAVIOR-1K, and real-world tasks demonstrate the effectiveness of MobiAgent. It outperforms π0.5-TA by 22.5 percentage points on BEHAVIOR-1K and enables robust recovery from execution failures. Through autonomous data recycling, success improves from 7.50% to 27.50% on RoboCasa and from 32.5% to 57.5% on Astribot S1.
THE CHALLENGE
A mobile manipulator must coordinate navigation and manipulation across an entire environment. Every subtask depends on the last succeeding; one dropped object can derail everything that follows.
MobiAgent closes the loop. The Inner Loop composes atomic skills and checks their outcomes during deployment. The Outer Loop builds a skill library from startup data, then uses verified deployment clips to refine the existing policies.
SEE IT IN ACTION
Navigate, grasp, pour, and return the bottle.
Pick up the bottle, carry it, and place it in the bin.
How the critic chooses to advance, retry, or replan.
Choose a chapter Original 4K video
THE FRAMEWORK
Deployment-time planning and recovery feed an offline loop for recursive policy self-improvement.
INNER LOOP · DEPLOYMENT
Compose the next subtask using observations, episode memory, and the atomic skill catalog.
Generate action chunks with a shared VLM trunk and skill-specific flow-matching experts.
Verify the visual outcome and return a verdict, follow-up action, and observable cues.
Visual feedback returns to the planner.
OUTER LOOP · TRAINING
Segment startup and deployment trajectories, then verify the captioned skill clips.
Cluster startup instructions into canonical skills, then map new rollout clips to the existing inventory.
Bootstrap skill experts, then refine the existing policies using curated deployment data.
Updated skills return to the deployment loop.
INSIDE THE SKILL EXECUTOR
The instruction selects an action expert. Visual and language understanding stay shared across the skill library.
Select an atomic skill
“Pick up the soda can.”
The picking expert handles grasping; the same expert can be reused for different objects and long-horizon tasks.
Choose a BEHAVIOR-style atomic skill to explore its expert and example instruction.
EXPERIMENTS
65.0%
Mean over four BEHAVIOR-1K tasks.
+22.5pp
65.0% compared with 42.5% for π0.5-TA.
32.557.5%
+25.0 percentage points over two training rounds.
ROBOCASA365 · COMPOSITE-SEEN
16 household tasks, with 2–15 subtasks each. A Franka arm on an Omron mobile base is evaluated with five trials per target scene; success is averaged across tasks and scenes.
Base π0.5 → best MobiAgent round
+20.00 percentage points at R3 and R5.
Early gains are followed by a plateau: R4–R6 range from 25.00% to 27.50%, ending at 26.25% in R6.
About the RoboCasa365 benchmark ↗| CaP-X | π0.5 (R0) | R1 | R2 | R3 | R4 | R5 | R6 |
|---|---|---|---|---|---|---|---|
| 5.00 | 7.50 | 18.75 | 21.25 | 27.50 | 25.00 | 27.50 | 26.25 |
Mean success across four BEHAVIOR-1K tasks, with 10 trials per task under matched evaluation conditions.
+22.5 percentage points over the task-specific baseline.
On Astribot S1, 100 demonstrations per task initialize the skill policies. Each update adds approximately 20 automatically curated deployment trajectories per task, without manual annotation.
mean success rate
+25.0 points from bootstrap
| Method | Push radio | Dispose trash | Stack storage | Fetch beer | Mean |
|---|---|---|---|---|---|
| π0.5-TA | 60 | 40 | 60 | 10 | 42.5 |
| MobiAgent | 70 | 60 | 70 | 60 | 65.0 |
COMPONENT ABLATIONS · BEHAVIOR-1K
Task success rate %
Global replanning adds 25.0 percentage points to mean success: 40.0% → 65.0%.
| Method | Push radio | Dispose trash | Stack storage | Fetch beer | Mean |
|---|---|---|---|---|---|
| Fixed task-specific sub-tasks (w/o Planner) | 50 | 0 | 0 | 30 | 20.0 |
| Single shared action head | 70 | 40 | 50 | 40 | 50.0 |
| Without Reflection Critic | 30 | 10 | 0 | 0 | 10.0 |
| Without global replanning | 70 | 40 | 30 | 20 | 40.0 |
| MobiAgent (Full) | 70 | 60 | 70 | 60 | 65.0 |
CITE THIS WORK
arXiv: 2610.03476 ↗
@misc{liu2026mobiagent,
title = {MobiAgent: Dual-Loop Recursive Policy Self-Improvement for Long-Horizon Mobile Manipulation},
author = {Chenzhi Liu and Yue Zhang and Jiehong Lin and Jianan Wang and Bo Wang and Zhongrui Wang and Xiaojuan Qi},
year = {2026},
archivePrefix = {arXiv},
eprint = {2610.03476},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.03476}
}