hierarchical replanning

Real-time robot replanning in dynamic environments incurs high computational overhead and relies heavily on explicit change detection and graph updates. Method: This paper proposes a novel incremental planning paradigm that eliminates the need for explicit reuse or update of historical paths. It decouples dynamic replanning into a sequence of independent, asymptotically optimal sampling-based planning problems—thereby avoiding dependence on obstacle change perception and dense graph reconstruction. The approach leverages almost-surely asymptotically optimal algorithms, including Effort-Informed Trees* (EIT*) and Asymptotically Optimal RRT-Connect (AORRTC), to balance rapid initial solution generation with continuous path refinement. Contribution/Results: Simulation results show that EIT*-generated paths achieve significantly shorter median lengths than those produced by mainstream reactive planners. Physical experiments on a robotic manipulator demonstrate AORRTC’s effectiveness and robustness in complex, dynamic task scenarios.

Addressing the challenge of real-time trajectory replanning for large-scale robot swarms in complex environments—requiring collision-free, deadlock-free, dynamically feasible, and computationally efficient solutions—this paper proposes a hierarchical cooperative framework. First, the workspace is partitioned spatially; then, conflict-free paths are computed in parallel within each partition; finally, distributed trajectory optimization, incorporating control feasibility constraints, ensures deadlock immunity and motion smoothness. The framework synergistically integrates centralized coordination with decentralized execution. To our knowledge, it is the first to achieve high-success-rate, real-time, deadlock-free replanning for swarms of over one hundred agents. Extensive simulations demonstrate real-time performance with 142 agents, while physical experiments on Crazyflie nano-quadcopters successfully deploy 24 robots. Compared to purely decentralized approaches, the method achieves significantly higher task success rates, with zero collisions and zero deadlocks throughout all evaluations.

This work addresses the challenge of efficiently replanning large-scale disturbances caused by physical couplings in industrial multi-agent systems under communication constraints and delays. The authors propose CASCADE, a novel mechanism that explicitly models communication range as an auditable and scalable coordination dimension. By decoupling a unified agent substrate from a scoped interaction layer, CASCADE enables dynamically triggered cascading coordination. Agents make local decisions conditioned on their roles using a shared knowledge base and employ lightweight contract primitives to expand coordination scope on demand. Evaluated in manufacturing and supply chain disruption scenarios, the approach achieves a superior trade-off among replanning quality, latency, and communication overhead, significantly enhancing robustness under uncertainty.

This work addresses the challenge of cascading conflicts in multi-agent cooperative scheduling, where delays of individual agents can propagate and disrupt global coordination. To mitigate this, the authors introduce the concept of temporal flexibility, which quantifies the maximum delay each agent can tolerate without violating the global schedule order. Building on this notion, they propose the FlexSIPP algorithm, which precomputes feasible alternative paths for potentially affected agents and dynamically exploits temporal slack through time-dependent search to enable efficient, cascade-free replanning. Empirical evaluation on the Dutch railway network demonstrates that the method generates robust, operationally compliant schedules within practical time limits, significantly enhancing the system’s resilience to disturbances.

This work addresses the limitations of existing large language model (LLM) agents, which typically employ fixed-granularity planning mechanisms that struggle to balance efficiency on simple tasks with the detailed reasoning required for complex ones. To overcome this, the paper introduces AdaPlan-H, a cognitively inspired adaptive hierarchical planning framework that, for the first time, integrates a progressive refinement strategy into LLM agents to enable dynamic, task-difficulty-aware adjustment of planning granularity. By synergistically combining hierarchical task decomposition, imitation learning, and capability enhancement, AdaPlan-H supports the adaptive generation and continuous optimization of planning hierarchies. Experimental results demonstrate that AdaPlan-H significantly improves success rates on multi-step complex tasks while effectively avoiding over-planning, thereby validating its efficiency and flexibility.

This work addresses the challenge of balancing efficiency and adaptability in large language model (LLM)-driven multi-agent coordination within dynamic environments. To this end, we propose SyncPlan, a framework that generates action chains for all agents through a single invocation of a centralized LLM, augmented with explicit synchronization primitives, a lightweight plan staleness detector, and an adaptive replanning mechanism. This integration enables highly efficient and robust long-horizon collaboration at minimal computational overhead. SyncPlan is the first to jointly incorporate explicit synchronization, deadlock detection, and planning optimization, leveraging both supervised fine-tuning and planning-oriented reinforcement learning. Evaluated on the Overcooked and Honor of Kings benchmarks, SyncPlan achieves state-of-the-art task success rates while consuming less than 0.05% of the runtime required by existing LLM-based coordination approaches.

This study addresses the high computational overhead and low deployment efficiency caused by conservative budgets in visual world model planning by proposing SufficientPlan, a training-free framework. The core innovations include Pairwise Sequential Budget Certification (PSBC), which dynamically identifies task-varying sufficient budgets through closed-loop evidence, and Static Context Reuse (SCR), which enables cross-iteration caching to eliminate redundant computation. Together, these mechanisms adaptively reduce search budgets without modifying pretrained models. Experiments demonstrate that the proposed approach significantly decreases search costs and inference latency across multiple backbones and control tasks while maintaining highly competitive control performance.

This work addresses the limitation of existing large language model–based code repair agents, which typically process tasks in isolation and struggle to reuse past repair experiences. To overcome this, the authors propose STAIR, a novel framework that introduces a hierarchical trajectory abstraction mechanism. STAIR constructs multi-granularity plan trees from historical repair trajectories, spanning from diagnostic actions to high-level strategies, and dynamically retrieves and adapts relevant nodes during new tasks to generate customized repair plans. Notably, this approach enables efficient cross-task and cross-agent experience transfer without modifying the target agent’s code. Experimental results on SWE-bench Verified demonstrate substantial performance gains: when integrated with the Lingxi agent, STAIR achieves Pass@1 scores of 81.2% (MiniMax M2.5) and 79.2% (GPT-5), and boosts mini-SWE-agent v2’s performance from 75.8% to 81.0%.

This study addresses the coordination failures between planning and execution agents in long-horizon tasks caused by inconsistent state cognition. To mitigate this issue, we propose ConPAct, a framework that localizes state conflicts through structured state assertions and programmatic contradiction detection. Furthermore, it integrates consistency-interactive fine-tuning to enable inference-time correction and collaborative training. Experimental results demonstrate that ConPAct improves task success rates on MiniGrid from 38.6% to 54.4% while significantly enhancing cross-environment generalization capabilities. By effectively aligning state representations across heterogeneous agents, this work establishes a novel paradigm for efficient multi-agent collaboration in complex sequential decision-making environments.

This study addresses the challenge of evaluating real-time motion planning in dynamic hazard fields by establishing a unified benchmark within rotating hazardous environments to systematically compare classical planning and learning-based paradigms. Through experiments employing classical planners, Proximal Policy Optimization (PPO) reinforcement learning, and dynamic obstacle simulation, this work reveals that environmental uncertainty is the predominant factor determining the effectiveness of a given planning paradigm. The results demonstrate that in stochastic dynamic environments, the PPO approach significantly outperforms classical methods in terms of computational latency, planning success rate, and path quality. These findings provide critical theoretical justification and empirical evidence for selecting appropriate motion planning paradigms for autonomous agents operating in complex scenarios.