ProLoop: Progress-Aware Closed-Loop Framework via Structured State Relations for Long-Horizon Robotic Task Planning

Runxin Zou; Qingmiao Liang; Jiatao Zhang; Xinyan Lv; Liangyu Qie; Meiyi Bu; Wei Song; Shiqiang Zhu

cs.ROOct 29, 2026

ProLoop: Progress-Aware Closed-Loop Framework via Structured State Relations for Long-Horizon Robotic Task Planning

Authors: Runxin Zou, Qingmiao Liang, Jiatao Zhang, Xinyan Lv, Liangyu Qie, Meiyi Bu, Wei Song, Shiqiang Zhu

Organizations: Zhejiang University, Hangzhou, China

Abstract

Long-horizon robotic tasks are commonly decomposed into semantic subgoals to simplify planning. However, subgoal progress in LLM-based planning is typically inferred implicitly from interaction histories. Inaccurate estimates can misguide execution and allow local errors to propagate. To address this, we propose ProLoop, a Progress-Aware Closed-Loop framework that derives explicit subgoal progress from structured relations between desired and execution states. A learned State-Relation Monitor assesses these relations, which are then aggregated into subgoal progress as ongoing, deviated, or completed. The resulting progress guides execution adaptation, enabling reliable progress-aware control throughout execution. Experiments on AgentBoard ALFWorld and a real-world robotic platform show that ProLoop improves success rate by 12.94% overall and by 15.46% on Hard tasks, while reducing environment interactions by 32.3%.

Figures & tables

Fig. 1: Compared with existing approaches, ProLoop explicitly derives subgoal progress from the relation between desired and execution states, distinguishes ongoing , deviated , and completed execution, and uses the resulting progress to support timely closed-loop adaptation.

Fig. 2: Overview of ProLoop. The LLM-Based Subgoal Planner generates the active subgoal and retention cues. Structured State Construction derives comparable desired and execution states, the State-Relation Monitor infers subgoal progress from their structured relations, and Progress-Guided Adaptation regulates continuation, targeted correction, or subgoal transition, with memory updated upon completion.

Fig. 3: Worked example of state-relation monitoring and progress derivation. For the active subgoal Heat the mug , desired conditions are compared with the corresponding execution state, and the resulting relations are aggregated to derive ongoing , deviated , or completed subgoal progress.

Fig. 4: Representative ALFWorld executions with ProLoop. Ongoing progress allows execution to continue without additional intervention, deviated progress triggers targeted correction, and completed progress enables subgoal transition and memory update.

Method	Overall (%)			Easy (%)		Hard (%)
Method	SR $\uparrow$	PR $\uparrow$	GA $\uparrow$	SR $\uparrow$	PR $\uparrow$	SR $\uparrow$	PR $\uparrow$
Model: Qwen3-14B
Plan-Only	$57.71{\scriptstyle\pm 2.46}$	$67.77{\scriptstyle\pm 1.42}$	$82.02{\scriptstyle\pm 1.29}$	$81.94{\scriptstyle\pm 1.96}$	$84.26{\scriptstyle\pm 1.73}$	$52.42{\scriptstyle\pm 3.09}$	$64.17{\scriptstyle\pm 1.98}$
Reason-Only	$83.58{\scriptstyle\pm 1.06}$	$91.21{\scriptstyle\pm 0.44}$	$93.18{\scriptstyle\pm 0.38}$	$94.44{\scriptstyle\pm 5.20}$	$95.37{\scriptstyle\pm 4.72}$	$81.21{\scriptstyle\pm 1.87}$	$90.30{\scriptstyle\pm 0.95}$
Reason-Hierarchical	$69.90{\scriptstyle\pm 1.41}$	$79.71{\scriptstyle\pm 1.41}$	$94.53{\scriptstyle\pm 0.06}$	$90.28{\scriptstyle\pm 1.96}$	$91.20{\scriptstyle\pm 0.65}$	$65.45{\scriptstyle\pm 1.96}$	$77.20{\scriptstyle\pm 1.75}$
Reason-Replan	$75.62{\scriptstyle\pm 2.54}$	$83.46{\scriptstyle\pm 1.56}$	$91.88{\scriptstyle\pm 1.08}$	$\mathbf{100.00}{\scriptstyle\pm 0.00}$	$\mathbf{100.00}{\scriptstyle\pm 0.00}$	$70.30{\scriptstyle\pm 3.09}$	$79.85{\scriptstyle\pm 1.90}$

TABLE I: Performance of ProLoop and baselines on the AgentBoard ALFWorld OOD split. Results for both backbones are reported as mean

\pm

standard deviation over three runs with different random seeds.

Method	Env. steps $\downarrow$	Tokens / episode $\downarrow$
Plan-Only	$20.61{\scriptstyle\pm 0.47}$	$31,676{\scriptstyle\pm 875}$
Reason-Only	$14.98{\scriptstyle\pm 0.23}$	$54,713{\scriptstyle\pm 2,629}$
Reason-Hierarchical	$17.25{\scriptstyle\pm 0.10}$	$35,950{\scriptstyle\pm 25}$
Reason-Replan	$17.29{\scriptstyle\pm 0.11}$	$\mathbf{8,422{\scriptstyle\pm 104}}$
Ours	$\mathbf{10.14{\scriptstyle\pm 0.13}}$	$25,191{\scriptstyle\pm 171}$

TABLE II: Environment interaction and token efficiency on the AgentBoard ALFWorld OOD split with Qwen3-14B. Results are reported as mean

\pm

standard deviation over three runs.

Fig. 5: Real-world execution of a breakfast preparation task with ProLoop on the dual-arm mobile robot.

Table 8

Robot Task Planning

Robot task planning turns goals and observations into executable actions while coordinating skills, motion, uncertainty, and safety.

See all 242 papers

ProAct-VLM: Pre-Failure Vision-Language Task Replanning with Continuous Perception Feedback
Sep 29, 2026Ahmed Nader Ahmed, Omar Moured, Mughni Irfan Mohammed Abdul +2Robot Policy Adaptation Long-Horizon Robotic Manipulation
Khalifa University, Abu Dhabi, United Arab Emirates. · Sereact GmbH, Stuttgart, Germany.
Metro-WM: Long-Horizon Latent Planning with Realisable Sub-Goals
Sep 25, 2026Royson Lee, Fady Rezk, Titouan Parcollet +2Long-Horizon Planning Goal-Conditioned Planning
Samsung AI, Cambridge, UK · University of Edinburgh, UK
RoboLongR: Causally Consistent Rewards Teach Robots to Progressively Master Long-Horizon Manipulation
Oct 24, 2026Zixuan Chen, Lin Cao, Fenggang Zhang +5Long-Horizon Robotic Manipulation Robot Policy Adaptation
Nanjing University · Southeast University · Fudan University

Sep 29, 2026J/K move · Enter open · S save

ProAct-VLM: Pre-Failure Vision-Language Task Replanning with Continuous Perception Feedback

Ahmed Nader Ahmed, Omar Moured, Mughni Irfan Mohammed Abdul, Muhayy Ud Din+1

Khalifa University, Abu Dhabi, United Arab Emirates. · Sereact GmbH, Stuttgart, Germany.

Robot Policy Adaptation Long-Horizon Robotic Manipulation Robot Task Planning VLMs for Robotics

Open paper AskPDF arXiv

Long-horizon robotic tasks are vulnerable to unexpected environmental changes that can render planned actions ineffective or unsafe. To address this, robots must detect such changes as they occur, interpret their impact, and adjust their actions accordingly. Traditional rule-based decision-making pipelines are brittle in open-world conditions, as they are hand-tuned for specific scenarios and lack generalization. Vision-Language Models (VLMs) offer a promising alternative as they combine broad world knowledge with unified visual--text reasoning, enabling them to generalize across diverse scenarios and generate accurate, grounded task plans. However, for effective deployment in dynamic real-world settings, VLMs must be embedded into frameworks capable of handling uncertainty and environmental changes. Existing frameworks broadly address this reactively, triggering replanning only after execution failures or post-task checks, risking failed actions. Some methods verify conditions before actions, but these discrete checks miss changes occurring during execution. To address this, we present ProAct-VLM, an adaptive, physically grounded task planning framework that integrates VLMs within a real-time perception--feedback loop. ProAct-VLM continuously monitors the environment and re-plans as soon as relevant changes are detected, enabling adaptation before failure occurs. Evaluations against multiple baselines and across different VLM backbones show that our framework improves both success rates and efficiency in dynamic, long-horizon manipulation tasks. Project page: https://github.com/moured/ProAct-VLM

Abstract

Robot Task Planning

Explore similar work

ProAct-VLM: Pre-Failure Vision-Language Task Replanning with Continuous Perception Feedback

Metro-WM: Long-Horizon Latent Planning with Realisable Sub-Goals

RoboLongR: Causally Consistent Rewards Teach Robots to Progressively Master Long-Horizon Manipulation