Lifelong LaCAM with Local Guidance for Lifelong MAPF

Abstract: Local guidance has recently proven to be a powerful driver of empirical performance in real-time, suboptimal multi-agent pathfinding (MAPF), improving the scalable configuration-based solver LaCAM. By injecting informative spatiotemporal cues around each agent, local guidance mitigates congestion, reduces waiting, and remains scalable enough even with tight time budgets, yielding state-of-the-art performance for one-shot MAPF. This study asks whether the same benefits can be lifted to the lifelong setting (LMAPF), where tasks arrive continuously and improvements in per-step plans can increase task completion throughput over long horizons. We propose LLLG, a Lifelong version of LaCAM enhanced with Local Guidance, which employs a receding-horizon windowed planning framework and warm-starts guidance from the previous solution at each timestep. Our method scales effectively, maintains high throughput even in compact, dense environments, and surpasses existing planners, thereby pushing the frontier of real-time, lifelong MAPF.

• The paper presents LLLG, a novel framework integrating receding-horizon windowed planning with local congestion-aware cost shaping and warm starting to improve throughput and reduce computational time.
• It employs rolling space-time A* for local guidance, mitigating near-term bottlenecks by dynamically balancing goal-directed costs with congestion penalties.
• Empirical results show up to 81% improved throughput and a 96% reduction in runtime, demonstrating LLLG's scalability and efficiency in dense, real-world multi-agent settings.

Lifelong LaCAM with Local Guidance (LLLG) for Lifelong MAPF: Technical Essay

Overview

The paper "Lifelong LaCAM with Local Guidance for Lifelong MAPF" (2605.16855) introduces LLLG, a receding-horizon, windowed planning framework that extends configuration-based search with local spatiotemporal guidance to tackle lifelong multi-agent pathfinding (LMAPF). The paper integrates windowed planning, local congestion-aware cost shaping, and incremental plan warm starting, resulting in a method that empirically attains marked improvements in throughput and computational efficiency in dense, operational-scale environments.

Problem Setting and Context

Lifelong MAPF formalizes real-world multi-robot coordination under online task arrival, requiring not just one-off collision-free solutions, but sustained, high-throughput operation over long time horizons on a static or dynamic undirected graph. Solutions must comply with strict per-timestep computational budgets, particularly in compact, high-density scenarios with hundreds to tens of thousands of agents. High agent densities exacerbate congestion, increasing the frequency and severity of joint conflicts, while online goal assignment introduces additional nonstationarity at each timestep.

Traditional MAPF techniques are ill-suited for lifelong settings: optimal or near-optimal solvers such as RHCR (Rolling Horizon Collision Resolution) and CBS (Conflict-Based Search) exhibit rapid performance degradation in dense or time-constrained settings; suboptimal planners (e.g., PIBT, LaCAM) scale better but suffer from severe suboptimality and fail to mitigate congestion. Guidance-based approaches using global, time-independent cost shaping provide some improvement but are ineffective for resolving fine-grained, spatiotemporal bottlenecks endemic to dense lifelong settings.

Methodological Contributions

The core technical contributions of LLLG are:

1. Receding-Horizon, Windowed Local Guidance: LLLG operates in a receding-horizon model: at each timestep, it replans over a finite window ($w_\Pi$), executes the first joint action, then repeats. This design enables immediate responsiveness to changing goals and congestion while amortizing planning cost.
2. Integration of Local Spatiotemporal Guidance: The method leverages and extends local guidance previously introduced for one-shot MAPF. For each agent, guidance paths are computed using rolling space-time A$^*$, balancing goal-directed cost with local congestion penalties informed by anticipated agent density in both space and time. This cost shaping directly biases the configuration generator to prefer actions likely to mitigate near-term bottlenecks, penalizing predicted high-traffic moves and excessive waiting.
3. Incremental Warm Starting with Prior Plans: Rather than reconstruct guidance from scratch each timestep, LLLG initializes each guidance instance by warm starting from the suffix of the previous windowed plan ($\Pi^{t-1}[2:w_\Phi]$), yielding an explicit, collision-free prediction of near-term joint behavior. This is shown empirically superior to inheriting only the prior timestep's local guidance, which provides a weaker, soft constraint on congestion.
4. Efficient, Anytime Refinement: Each windowed plan can be improved within residual per-step time using LaCAM$^*$ or LNS; however, the paper finds that additional sub-window refinement often has marginal benefit, as full-horizon plan structure quickly becomes invalid due to dynamic goals.
5. Integration of Hindrance: The method further enhances local cost construction with a hindrance estimate, quantifying the potential of an agent's tentative move to block others in the subsequent timestep, leading to significant improvement with negligible runtime impact.

Empirical Evaluation

Strong numerical results demonstrate that LLLG achieves:

• Superior Throughput:

Across a variety of high-density map benchmarks (e.g., 1,000 agents in grid environments), LLLG consistently attains higher average task completion rates than state-of-the-art baselines. On challenging instances with bottlenecks (such as ht_chantry), LLLG outperforms RHCR in throughput by up to 81% while operating under a 96% reduction in runtime versus RHCR.

• Computational Efficiency and Scalability:

LLLG maintains sub-second per-timestep runtimes even for dense cases involving up to 10,000 agents, demonstrating practical deployment viability. Unlike RHCR and other branch-and-replan frameworks, runtime does not explode with agent count, and solution feasibility is consistently maintained.

• Congestion Mitigation:

Visualizations show that while previous methods concentrate waiting and agent conflicts at map centers/bottlenecks, LLLG achieves a more uniform spatiotemporal distribution of agent stops and dynamically diffuses congestion.

• Effectiveness of Warm Start and Guidance Window:

Warm starting from the previous windowed plan is shown vital for throughput gains, even at identical computational cost. There exists a broad, diminishing-returns plateau as planning horizon and guidance window increase; excessively large windows do not further improve throughput due to unpredictability in future task assignments.

• Resilience to Pathological Map Structures:

Certain failure cases remain for environments with long, narrow corridors where local guidance is unable to anticipate interactions outside its window.

Limitations and Implications

The investigation reveals several key points for practical and theoretical development:

• Window Size Versus Realism:

Windowed planning must balance plan quality with responsiveness: longer windows resolve near-term conflicts better but offer sharply reduced marginal benefit as future goals are unknown. The practical limit for window/horizon size is dictated by the stochasticity of goal assignment and real-time constraints.

• Role of Local versus Global Guidance:

Local spatiotemporal guidance significantly outperforms time-independent, globally-derived heuristics in dense regimes, suggesting that agent-centric, short-range prediction is sufficient for flow-level coordination in practical settings.

• Anytime Refinement Disconnect:

Efforts to improve windowed plan cost through branch-and-bound or LNS yield marginal or negative impact on long-term throughput, highlighting a cost-objective mismatch between per-window planning and the true infinite-horizon LMAPF metric.

• Scalability:

LLLG provides a viable algorithmic template for warehouse-scale multi-agent deployments, notably in logistics, automated storage/retrieval, and large fleet management.

• Future Directions:

Further improvements could arise from integration of information about future goal distributions or known task pipelines, as well as learning-based augmentation to better predict high-congestion events beyond the immediate planning window.

Conclusion

LLLG achieves a new tradeoff frontier for LMAPF, combining scalable, configuration-based planning with efficient, spatiotemporally-local cost shaping and incremental plan reuse. It empirically surpasses existing solvers in both throughput and runtime under dense and large-scale conditions, with notable robustness to map topology and agent density. The approach demonstrates the centrality of local online guidance, joint-space plan reuse, and efficient congestion prediction for scalable lifelong multi-agent pathfinding in realistic robotic systems.