PaperlayerView exact revision on arXiv

Source-linked AI summary

MobiAgent: Dual-Loop Recursive Policy Self-Improvement for Long-Horizon Mobile Manipulation

Chenzhi Liu, Yue Zhang, Jiehong Lin, Jianan Wang, Bo Wang, Zhongrui Wang, Xiaojuan Qi

arXiv:2610.03476v1cs.ROcs.AI

TL;DR

Long-horizon mobile manipulation is challenged by compounding execution errors and capacity interference, while existing systems have limited hierarchical reasoning, flexible replanning, and continuous learning. MobiAgent combines composable atomic skills in a deployment-time Inner Loop with an offline Outer Loop for autonomous data curation and policy updates. It reports gains on BEHAVIOR-1K, RoboCasa, and real-world Astribot S1 tasks.

  • Problem

    Long-horizon mobile manipulation requires interleaving navigation and precise manipulation, but compounding errors and capacity interference challenge joint control, while existing agents have limited hierarchical reasoning and flexible replanning.

    Source: §1 · p. 1

  • Method

    MobiAgent separates deployment-time execution from offline policy evolution, composing atomic skills with a receding-horizon Inner Loop and using curated rollouts to update skill policies in an Outer Loop.

    Source: §1 · pp. 2–3

  • Results

    MobiAgent exceeds π0.5-TA by 22.5 percentage points on BEHAVIOR-1K, while success rises from 7.50% to 27.50% on RoboCasa and from 32.5% to 57.5% on Astribot S1.

    Source: §5 · p. 9

  • Takeaways & Limitations

    The results support deployment-to-training data recycling as a route to annotation-free policy self-improvement across simulated and real-world tasks.

    Source: §1 · p. 3, §5 · p. 9

    Adding new skills to the shared VLA backbone may interfere with previously learned behaviors.

    Source: §5 · p. 9

Abstract

from arXiv · show

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 $\pi_{0.5}$-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.

1 Introduction

Long-horizon mobile manipulation combines compounding execution errors and capacity interference, while existing reactive and hierarchical systems have difficulty with long-term reasoning, flexible replanning, and continued learning. MobiAgent addresses these gaps with composable atomic skills, a deployment-time Inner Loop, and an offline Outer Loop; evaluations report benchmark gains and improved success after data recycling.

1 Introduction

  • Long-horizon tasks require interleaving navigation and precise prehension, but compounding errors and capacity interference challenge joint locomotion and arm control.Recent VLA policies are effective on short-horizon tasks but lack hierarchical reasoning and long-term context for closed-loop replanning.Source: §1 · pp. 1–2
  • Existing hierarchical agents often constrain planning to predefined sub-task policies, make failure-triggered replanning cumbersome, and lack comprehensive data curation for recursive self-improvement.These limitations are especially consequential for mobile recovery, which requires continuous spatial reasoning before interaction.Source: §1 · p. 2
  • The Inner Loop dynamically composes atomic skills in a receding-horizon flow, using shared-backbone, skill-specific flow-matching experts and visual verification for flexible replanning.Sharing action experts supports reuse while avoiding capacity interference; the Reflection Critic manages spatial reasoning for error recovery.Source: §1 · p. 3
  • The Outer Loop segments and verifies demonstrations and deployment rollouts, clusters instructions into an atomic-skill vocabulary, and uses curated data to instantiate and fine-tune policies.This automated curation and skill-library evolution requires no manual data curation.Source: §1 · p. 3
  • 65.0% mean success on BEHAVIOR-1K exceeds π0.5-TA by 22.5 percentage points; updates raise RoboCasa success from 7.50% to 27.50% and Astribot S1 success from 32.5% to 57.5%.The evaluations also report critic-driven recovery from object drops and pose misalignments.Source: §1 · p. 3

2 Related Work

Prior work includes short-horizon VLA policies, hierarchical plan-execute-verify agents, and mobile manipulation systems focused on reactive control or navigation-manipulation hand-offs. MobiAgent is positioned as combining skill-conditioned vision-based control, adaptive replanning, and deployment-driven refinement.

Vision-language-action (VLA) policies for short-horizon manipulation

  • VLA approaches generate short-horizon actions either as discretized action tokens or, in π0 and π0.5, as continuous trajectories through flow matching.RT-2 and OpenVLA exemplify the token approach; π0 and π0.5 exemplify flow matching.Source: §2 · p. 3

Hierarchical agentic frameworks for long-horizon tasks

  • Hierarchical agents often use fixed skill repertoires and rigid recovery, whereas MobiAgent uses a unified skill-conditioned VLA, skill-level verification, task-level replanning, and rollout-based refinement.RoboClaw, for example, requires explicitly paired forward and reset policies.Source: §2 · p. 3

Mobile manipulation systems

  • Mobile manipulation systems and VLA extensions excel at reactive control and navigation-manipulation hand-offs but lack long-horizon closed-loop replanning, online verification, and deployment-driven refinement.π0.5-TA uses isolated task-specific models whose time-indexed stages make distinct behaviors share one prediction head.Source: §2 · p. 3

3 MobiAgent: A Dual-Loop Agentic System for Mobile Manipulation

MobiAgent separates deployment-time execution from offline policy evolution. Its inner loop repeatedly plans, executes, and verifies, while the outer loop curates rollouts and updates the skill library.

3.1 The Inner Loop: Closed-Loop Agentic Execution

  • The inner loop recurrently plans, executes, and evaluates skills using current observations and recent failures, continuing until task success or a fixed budget is exhausted.Source: §3.1 · p. 5

3.1.1 Task Planner: Dynamic Atomic Skill Composition

  • The Task Planner dynamically composes fine-grained skills at decision points, replanning after skill completion or failure to adapt to environmental changes.Source: §3.1.1 · p. 5

3.1.2 Skill Executor: Skill-Specific Action Generation

  • Skill-specific flow-matching experts share a VLM trunk, supporting reusable policies while specializing behavioral modes to mitigate capacity interference.Source: §3.1.2 · p. 5

3.1.3 Reflection Critic: Continuous Evaluation and Feedback

  • The Reflection Critic assesses task progress from sampled observations and returns a status, follow-up recommendation, and observable cues for auditing.Source: §3.1.3 · pp. 5–6

3.2 The Outer Loop: Automated Agentic Training

  • The outer loop bootstraps from startup data, then processes deployment rollouts offline to curate clips, formalize skills, and instantiate or update policies.Source: §3.2 · p. 6

3.2.1 Data Curator: VLM-Powered Trajectory Segmenting

  • A video-capable VLM segments trajectories and captions clips, while a separate VLM verifies continuity, skill adjacency, and object-instruction consistency.Source: §3.2.1 · p. 6

3.2.2 Skill Generator: Atomic Skill Clustering

  • The Skill Generator clusters startup instructions, names and standardizes canonical skills, then assigns rollout clips to the existing vocabulary.Source: §3.2.2 · pp. 6–7

3.2.3 Skill Trainer: Continuous Policy Evolution

  • The Skill Trainer bootstraps dedicated experts from canonical skill clusters using behavior cloning, then fine-tunes existing policies on curated rollouts.Source: §3.2.3 · p. 7

4 Experiments

Experiments evaluate MobiAgent on RoboCasa, BEHAVIOR-1K, and Astribot S1, reporting benchmark results, component ablations, and recovery behavior. Results include gains from deployment-data updates and critic-driven global replanning.

4.1 Main Results

  • 27.50% success in RoboCasa Round 3, up from 18.75% in Round 1 and the base π0.5 policy’s 7.50%, before stabilizing at 25.00–27.50%.The reported evaluation covers 16 composite household tasks with two to fifteen subtasks.Source: §4.1 · p. 7
  • 57.5% mean success on Astribot S1 after two updates, up from 32.5% at bootstrap and 10.0% for π0.5.The Outer Loop automatically curates approximately 20 deployment trajectories per task after each round.Source: Table 3 · p. 8, §4.1 · p. 8
  • Under matched conditions with 10 trials per task, MobiAgent achieves 65.0% mean success, versus π0.5-TA’s 42.5%, a 22.5-percentage-point advantage.The gain reaches 50 points on the longest fetch-beer task.Source: §4.1 · p. 8

4.2 Ablation Studies

  • Replacing shared, task-agnostic atomic skills with fixed task-specific sub-task models reduces mean success from 65.0% to 20.0%.With shared atomic skills, semantically identical behaviors such as picking reuse one policy across tasks.Source: §4.2 · p. 8
  • Replacing skill-specific action experts with one action head lowers mean success from 65.0% to 42.5%.The authors associate specialized experts with mitigating interference among navigation, grasping, placement, and articulated-object interaction.Source: §4.2 · p. 8
  • Removing the Reflection Critic lowers mean success from 65.0% to 10.0%, as unverified skill outcomes let execution errors propagate.Without visual outcome verification, the system cannot reliably tell whether an active skill completed or failed.Source: §4.2 · p. 8

4.3 Robustness and Error Recovery

  • Global replanning raises mean success from 40.0% to 65.0%; 39% of successful episodes replan, averaging 0.95 replans under a three-replan budget.It updates the skill sequence for the observed post-failure scene, including object retrieval and base-pose correction before retrying a grasp.Source: §4.3 · p. 9, Figure 4 · p. 9

5 Conclusion and Limitations

MobiAgent combines dynamic atomic-skill execution with an offline loop for data curation and policy improvement. The conclusion reports gains across benchmarks and real-world tasks, while noting limitations in skill expansion, recovery, and evaluation scope.

5 Conclusion and Limitations

  • MobiAgent exceeds π0.5-TA by 22.5 percentage points on BEHAVIOR-1K, raises RoboCasa success from 7.50% to 27.50%, and increases Astribot S1 success from 32.5% to 57.5%.The real-world gain follows two policy-update rounds.Source: §5 · p. 9
  • Adding skills may interfere with previously learned behaviors; recovery also faces pose-correction churn, inaccurate placement, and critic drift under transient occlusions, while broader evaluations are needed.The authors suggest parameter-efficient modules and continual-learning strategies for skill expansion, and geometric perception and future–reality verification for recovery.Source: §5 · p. 9

A LLM/VLM Configurations and Prompt Design

  • GPT-5.4 powers most modules, while Gemini-3.1-Pro handles trajectory segmentation because it accepts continuous video input.Uniformly sampled frames degrade temporal boundary localization, according to the authors.Source: A LLM/VLM Configurations and Prompt… · p. 15
  • Task-agnostic prompts use canonical skills and verb-frame templates; task, room, and object context enters through the goal, history, and observation fields.The same prompts are reused across tasks, and constrained decoding enforces output schemas.Source: A LLM/VLM Configurations and Prompt… · p. 15

A.1 Task Planner

  • The Task Planner selects one atomic next step from the current goal, history, observations, and robot state, replanning after completion or failure.Its rules require checking prerequisites, avoiding redundant navigation, and using the post-failure scene to choose a supported next step.Source: §A.1 · pp. 15–16

A.2 Reflection Critic

  • The Reflection Critic judges only the active subtask from recent camera frames and proprioception, requiring positive evidence for completion and uncertainty-aware checks under occlusion.It distinguishes incomplete progress from confirmed deviation and recommends retry or replanning accordingly.Source: §A.2 · pp. 16–17

A.3 Data Curator - Segmenting

  • The Data Curator segments recordings into contiguous atomic clips, separating travel from state changes and splitting carried-object acquisition and placement into distinct clips.It labels state changes only when observed and forbids prep-only labels.Source: §A.3 · p. 17

A.4 Data Curator - Verification

  • The verifier accepts a segmentation only when canonical skills, plausible targets, and their ordered sequence are coherent for the goal; ambiguous or malformed sequences are rejected.The Skill Generator clusters free-form descriptions by action intent, assigns reusable canonical names, and folds preparation and completion phases into supported state-changing skills.Source: §A.4 · pp. 17–18, §A.5 · p. 18

B.2 Evaluated Tasks and Subtask Sequences

The supplementary sections describe evaluated task sequences, per-skill policy training, and the real-robot evaluation setup.

B.2 Evaluated Tasks and Subtask Sequences

  • BEHAVIOR-1K task instructions are passed unchanged to the Task Planner, and representative subtask sequences use canonical skills such as move to, pick up, and place.The sequences are described as structurally representative because segmentation is consistent across episodes.Source: §B.2 · p. 19

B.3 Skill Executor Policy Training

  • The Skill Executor is a single π0.5 model with a shared PaliGemma backbone and a distinct gemma 300m action expert per canonical skill, fine-tuned on skill-specific clips.The six BEHAVIOR-1K skills are jointly fine-tuned with batch shares proportional to dataset sizes; RoboCasa also jointly fine-tunes a shared backbone and six experts.Source: §B.3 · p. 19

C.1 Real-Robot Environment and Embodiment

  • Astribot S1 evaluation uses 10 trials per task and method under fixed scene resets; Figure S1 shows time-ordered rollouts for four tasks with executed skills labeled beneath frames.The platform uses head, wrist, and chest cameras, while demonstrations are collected through VR teleoperation and processed by the same curation pipeline as simulation.Source: §C.1 · pp. 19–20, Figure S1 · p. 21

C.2 Real-Robot Evaluated Tasks and Subtask Sequences

The real-robot evaluation uses unchanged task instructions and per-skill training sequences, with a shared-backbone executor fine-tuned on curated clips.

C.2 Real-Robot Evaluated Tasks and Subtask Sequences

  • Real-robot task instructions are passed unchanged to the Task Planner, and training sequences use canonical skills including move to, pick up, place, and pour.Each listed sequence line corresponds to a segment tagged with a canonical skill.Source: §C.2 · p. 21

C.3 Skill Executor Policy Training

  • The Astribot S1 executor is one π0.5 model with a shared PaliGemma backbone and five skill-specific experts, trained on combined trash and pour-blue clips.Runs freeze the vision backbone, use a per-timestamp action normalizer, and fit on 4× A100 GPUs.Source: §C.3 · p. 21

C.4 Additional Real-Robot Evaluation on Unitree R1D

MobiAgent is also evaluated on a multi-stage home-clean task with the Unitree R1D.

C.4 Additional Real-Robot Evaluation on Unitree R1D

  • 20.0% task success on Unitree R1D home-clean compares with 0.0% for the base π0.5 policy.The task includes door opening, bottle collection and disposal, and table cleaning.Source: §C.4 · pp. 21–22

D Additional Quantitative Results and Analysis

This section presents three quantitative analyses: data-curation wall-clock cost versus manual annotation, per-episode runtime versus an end-to-end baseline, and per-subtask success rates.

D.1 Data-collection Wall-clock Cost versus Manual Annotation

D.2 Per-episode Runtime

D.3 Per-subtask Success Rate

D.1 Data-collection Wall-clock Cost versus Manual Annotation

The analyses compare MobiAgent’s data-curation cost with manual annotation, examine runtime overhead, and report per-subtask success rates across simulation and real-world deployments.

D.1 Data-collection Wall-clock Cost versus Manual Annotation

  • At 800 clips, MobiAgent’s data processing takes ∼0.82 h versus ∼28.9 h for manual annotation, a ∼35× speedup.Source: §D.1 · p. 22, Figure S3 · p. 23

D.2 Per-episode Runtime

  • MobiAgent’s per-episode runtime includes physical action execution and VLM API time, with API overhead the primary bottleneck.The runtime comparison covers successful rollouts; π0.5-TA operates end-to-end without API calls.Source: §D.2 · pp. 22–23, Table S8 · p. 23

D.3 Per-subtask Success Rate

  • Per-subtask success rates by canonical skill are reported for simulation and real-robot instances successfully reached during execution.The analysis helps localize where residual failures concentrate.Source: §D.3 · p. 24, Table S9 · p. 24

E Failure Modes and Analysis

The failure analysis identifies critic false positives and bimanual coordination as operational bottlenecks, alongside handoff-pose and perception failures in its broader taxonomy.

E Failure Modes and Analysis

  • The analysis classifies execution errors into four bottlenecks: critic false positives, bimanual coordination, inter-skill handoff poses, and low-level perception.Source: E Failure Modes and Analysis · p. 24

E.1 FM1: Reflection Critic False Positives

  • A false-positive completion verdict can advance an out-of-range approach to an impossible pickup, exhausting retries before replanning detects the stalled task.The critic can later detect the downstream pickup failure and trigger replanning.Source: §E.1 · p. 24, Figure S4 · p. 25

E.2 FM2: Bimanual-Coordination Failures

  • Because one control head drives both arms without conditioning on the partner arm’s state, tightly coupled motions can knock over targets or cause poor synchronization.The failures include bottle retrieval and radio handling.Source: §E.2 · p. 24, Figure S5 · p. 25
  • In stack-storage, a one-gripper hold can tilt the large container and derail carrying and stacking; arm-specific descriptions and targeted training data are proposed for future work.Source: §E.2 · p. 25, Figure S5 · p. 25

E.3 FM3: Inter-skill Transition and Pose Mismatch

The analysis examines failures caused by inter-skill handoff poses and low-level perception, then compares MobiAgent with CaP-X on tasks within CaP-X’s navigate-and-grasp repertoire.

E.3 FM3: Inter-skill Transition and Pose Mismatch

  • An upstream skill can achieve its goal yet leave an out-of-distribution pose that the next skill inherits, causing placement failure despite no explicit upstream error.In dispose-trash, pickup succeeds but the misaligned handoff prevents placing the can over the bin opening.Source: §E.3 · p. 26, Figure S6 · p. 26

E.4 FM4: Low-level Perceptual Limits

  • Spatial and metric perception errors can place a bottle outside the bin, while difficulty localizing a clear glass can lead to pouring beside it.These real-robot failures are attributed to sensing and representation limits of the current action expert.Source: §E.4 · p. 26, Figure S7 · p. 27

F.1 Comparison with CaP-X

  • On CaP-X’s navigate-and-grasp task subset, MobiAgent navigates and grasps reliably and can continue with press or place steps that CaP-X cannot attempt.The comparison uses 10 trials per task on held-out test scenes; navigation and task success are reported separately.Source: §F.1 · p. 27, Table S10 · p. 27

F.2 Comparison with RoboClaw

MobiAgent and RoboClaw differ in both task decomposition and replanning: MobiAgent uses atomic skills and replans at every step, while RoboClaw uses coarse composites and a static plan with conditional recovery. MobiAgent’s fixed skill catalogue also yields more consistent decompositions across paraphrases, though the architectural mismatch precludes a direct quantitative comparison.

F.2 Comparison with RoboClaw

  • RoboClaw’s coarse composite subtasks bundle several MobiAgent atomic skills, so its plans cannot be routed directly to MobiAgent’s skill-conditioned experts.The authors report this architectural divergence rather than a direct head-to-head quantitative comparison.Source: §F.2 · pp. 27–28, Table S11 · p. 28
  • MobiAgent re-invokes its Task Planner at every step, while RoboClaw executes a full upfront plan unchanged unless a subtask’s success check fails.The comparison characterizes continuous replanning as the architectural driver of MobiAgent’s superior robustness.Source: §F.2 · p. 28, Table S12 · p. 29
  • Across paraphrases, MobiAgent’s atomic steps remain nearly constant in number and order, while RoboClaw’s free-form subtasks vary widely in count and granularity, especially on dispose-trash.The evaluation holds the task goal fixed across three paraphrases and uses the same GPT-5.4 backbone and starting frame for both planners.Source: §F.2 · p. 28, Table S13 · p. 29, Table S14 · p. 30
  • MobiAgent’s fixed catalogue constrains subtask granularity and keeps expert prompts short and in-distribution; RoboClaw’s free-form prompts can produce verbose, inconsistent strings.Paraphrasing changes MobiAgent’s object wording, not the overall step count, aside from a redundant per-bottle cycle for one fetch-beer wording.Source: §F.2 · pp. 28–29

Related papers

Loading 2610.03476v1…