LaCAM: Lazy Constraint MAPF Framework

• LaCAM is a search-based multi-agent pathfinding framework that lazily generates successor configurations through incremental constraint accumulation.
• It employs a two-level search architecture with high-level configuration exploration and low-level lazy constraint generation, effectively mitigating combinatorial explosion.
• Variants such as LaCAM*, Real-Time LaCAM, and db-LaCAM extend its capabilities to anytime optimization, real-time planning, and kinodynamic motion planning.

LaCAM is a search-based algorithmic framework for Multi-Agent Pathfinding (MAPF) that searches over joint agent configurations while avoiding explicit enumeration of the exponentially large successor set by generating successors lazily under incrementally accumulated constraints. In its original formulation, LaCAM was introduced as a complete, sub-optimal MAPF solver that remains practical for hundreds of agents or more; later work developed an anytime eventually optimal variant, a provably complete real-time version, guidance-enhanced and lifelong planners, and kinodynamic extensions for multi-robot motion planning and warehouse AGVs (Okumura, 2022, Okumura, 2023, Liang et al., 8 Apr 2025, Moldagalieva et al., 7 Dec 2025).

1. Problem formulation and base architecture

LaCAM is defined on the classical MAPF model over an undirected graph $G=(V,E)$ with agents $A=\{1,\dots,n\}$, distinct start vertices $\mathcal S=(s_1,\dots,s_n)\in V^n$, and distinct goal vertices $\mathcal G=(g_1,\dots,g_n)\in V^n$. A configuration is a joint state $Q=(q_1,\dots,q_n)\in V^n$, and a valid transition requires that each agent either waits or moves to an adjacent vertex without vertex or edge collisions. The original paper measures solution quality by the sum-of-costs,

$\mathrm{SoC}=\sum_{i=1}^n T_i,$

where $T_i$ is the first time at which agent $i$ reaches and stays at its goal (Okumura, 2022).

The original formulation presents LaCAM as a two-level search. At the high level, it explores configurations $Q\in V^n$. At the low level, it lazily enumerates partial constraints that specify where selected agents must go in the next timestep. A high-level node stores a configuration, a parent pointer for backtracking, an agent ordering $\sigma$, and a low-level queue or tree of partial constraints. A low-level constraint-node has the form $A=\{1,\dots,n\}$0, meaning that agent $A=\{1,\dots,n\}$1 must move to vertex $A=\{1,\dots,n\}$2 at the next global step. When a constraint-node is expanded, the next agent in $A=\{1,\dots,n\}$3 is assigned one of its admissible next vertices, and the resulting child constraints are queued lazily rather than fully materialized at once (Okumura, 2022).

The key operational component is the configuration generator. Instead of enumerating all joint moves, LaCAM calls a fast one-step subroutine—typically PIBT, or Priority Inheritance with Backtracking—to propose one collision-free successor configuration that respects the currently active constraints. This separation between high-level configuration search and low-level constrained successor generation is the main reason the algorithm avoids the immediate $A=\{1,\dots,n\}$4 blow-up of explicit successor enumeration. In later expositions, the same idea is described as a depth-first exploration of the joint-agent configuration graph with “just-in-time” constraint additions, emphasizing the lazy and incremental nature of the search (Liang et al., 8 Apr 2025).

2. Constraint semantics and completeness

The central mechanism in LaCAM is incremental constraint accumulation. In the original presentation, the low-level search lazily builds sets of positive next-step commitments such as $A=\{1,\dots,n\}$5, and every complete assignment of one next-step destination per agent can, in principle, force a distinct neighbor configuration. Because these assignments are generated one by one, LaCAM pays only for the successors it actually attempts to realize, rather than for all neighbors simultaneously (Okumura, 2022).

The Real-Time LaCAM exposition recasts the same search style in a more explicit action-constraint language. There, LaCAM is “best understood as a depth-first exploration of the joint-agent configuration graph” in which a growing set of action-constraints $A=\{1,\dots,n\}$6 forbids particular local moves. A single constraint is written $A=\{1,\dots,n\}$7, meaning that agent $A=\{1,\dots,n\}$8 is not allowed to take action $A=\{1,\dots,n\}$9. The DFS always invokes a one-step configuration generator to return exactly one valid successor under the current $\mathcal S=(s_1,\dots,s_n)\in V^n$0. If the successor has already been visited or becomes a dead end, LaCAM backjumps to the offending ancestor, adds one new constraint, and forces a different successor on the next attempt (Liang et al., 8 Apr 2025).

This view yields a concise completeness argument. Let $\mathcal S=(s_1,\dots,s_n)\in V^n$1 be the revisit count of configuration $\mathcal S=(s_1,\dots,s_n)\in V^n$2, and define

$\mathcal S=(s_1,\dots,s_n)\in V^n$3

Since each revisit adds at least one new constraint on one agent’s action out of its $\mathcal S=(s_1,\dots,s_n)\in V^n$4 possible moves, after at most

$\mathcal S=(s_1,\dots,s_n)\in V^n$5

constraints at $\mathcal S=(s_1,\dots,s_n)\in V^n$6, every neighbor will have been forced once. The Real-Time LaCAM paper states the corresponding theorem as follows: if there exists any collision-free path from the start configuration to the goal configuration, LaCAM’s lazy DFS, by accumulating constraints on every repeated visitation, will find it in finite time. The original LaCAM paper gives the same high-level conclusion from the two-level perspective: the low-level BFS eventually enumerates all relevant constraint-sets, so the high-level search eventually generates all reachable configurations and either finds a solution or exhausts the finite configuration graph (Okumura, 2022, Liang et al., 8 Apr 2025).

A plausible implication is that the literature provides two complementary but compatible interpretations of the same core design: one in terms of lazy positive next-step commitments, and another in terms of revisit-driven forbidden-action accumulation.

3. Anytime optimization and search engineering

LaCAM$\mathcal S=(s_1,\dots,s_n)\in V^n$7 extends LaCAM into an anytime algorithm that continues searching after the first solution and eventually converges to an optimum, provided that solution costs are accumulated transition costs. The core search state augments each configuration node with a path cost $\mathcal S=(s_1,\dots,s_n)\in V^n$8, an admissible heuristic $\mathcal S=(s_1,\dots,s_n)\in V^n$9, and an evaluation $\mathcal G=(g_1,\dots,g_n)\in V^n$0. When a goal configuration is found, it becomes the incumbent $\mathcal G=(g_1,\dots,g_n)\in V^n$1, but the search continues. If a previously seen configuration is reached by a cheaper path, LaCAM$\mathcal G=(g_1,\dots,g_n)\in V^n$2 performs a local Dijkstra-style repair or relaxation over the discovered neighbor graph, rewiring parents and propagating improved $\mathcal G=(g_1,\dots,g_n)\in V^n$3-values. Under true additive costs, this preserves completeness and yields eventual optimality when the $\mathcal G=(g_1,\dots,g_n)\in V^n$4-search is allowed to finish (Okumura, 2023, Okumura, 2023).

A commonly emphasized additive objective in the LaCAM$\mathcal G=(g_1,\dots,g_n)\in V^n$5 literature is the sum-of-loss,

$\mathcal G=(g_1,\dots,g_n)\in V^n$6

which is particularly convenient for Dijkstra-style rewiring and anytime refinement. The empirical motivation for LaCAM$\mathcal G=(g_1,\dots,g_n)\in V^n$7 is that baseline LaCAM quickly finds feasible solutions, but initial solution quality may be far from optimal and convergence can be slow (Okumura, 2023).

The engineering literature introduced five orthogonal quality-improvement techniques while preserving eventual optimality: non-deterministic node extraction, Space Utilization Optimization (SUO), Monte-Carlo Configuration Generation (MCCG), dynamic incorporation of alternative solutions, and recursive calls of LaCAM$\mathcal G=(g_1,\dots,g_n)\in V^n$8 itself as a refiner. In the 30 s benchmark summary over 800 instances, the original LaCAM$\mathcal G=(g_1,\dots,g_n)\in V^n$9 solved all 800 instances with average normalized cost $Q=(q_1,\dots,q_n)\in V^n$0, whereas the enhanced version solved $Q=(q_1,\dots,q_n)\in V^n$1 with average cost $Q=(q_1,\dots,q_n)\in V^n$2, corresponding to an $Q=(q_1,\dots,q_n)\in V^n$3 reduction (Okumura, 2023).

Later work replaced expensive static guide-path preprocessing with a dynamic Lightweight Traffic Map (LTM). The LTM is a directed weighted graph over workspace edges whose weights are updated online from PIBT’s committed actions, blocked actions, and propagated wait counts. PIBT then ranks moves using weighted distances $Q=(q_1,\dots,q_n)\in V^n$4 rather than unweighted shortest-path distances. This directly addresses two reported drawbacks of prior guidance-path methods—high offline cost and static bias—and the resulting LaCAM$Q=(q_1,\dots,q_n)\in V^n$5+LTM achieves higher solution quality than state-of-the-art guidance-path approaches, including $Q=(q_1,\dots,q_n)\in V^n$6 lower loss ratio in dense one-shot MAPF scenarios and faster anytime convergence (Shen et al., 9 Mar 2026).

4. Real-time, guided, and lifelong planning

Real-Time LaCAM interleaves LaCAM’s DFS construction with online execution. Instead of waiting for a full-horizon solution, it maintains a single global DFS tree $Q=(q_1,\dots,q_n)\in V^n$7 and constraint set $Q=(q_1,\dots,q_n)\in V^n$8, resumes the DFS for at most a cutoff time $Q=(q_1,\dots,q_n)\in V^n$9 milliseconds at each iteration, backtracks from the stopping point to the current configuration $\mathrm{SoC}=\sum_{i=1}^n T_i,$0 to extract the next move $\mathrm{SoC}=\sum_{i=1}^n T_i,$1, executes that move, and then reroots $\mathrm{SoC}=\sum_{i=1}^n T_i,$2 at $\mathrm{SoC}=\sum_{i=1}^n T_i,$3 by reversing the parent pointer on that edge. Because the same search tree is eventually built, Real-Time LaCAM has the same success rate—$\mathrm{SoC}=\sum_{i=1}^n T_i,$4 on solvable instances—as full-horizon LaCAM under any per-iteration cutoff $\mathrm{SoC}=\sum_{i=1}^n T_i,$5, and its total planning time is the full-horizon runtime plus negligible rerooting and backtracking overhead. The paper characterizes this as the first real-time MAPF method with provable completeness guarantees (Liang et al., 8 Apr 2025).

A distinct line of work improves empirical performance by injecting local spatiotemporal guidance into LaCAM’s configuration generator. At each LaCAM step, GuidanceConstruction solves $\mathrm{SoC}=\sum_{i=1}^n T_i,$6 decoupled $\mathrm{SoC}=\sum_{i=1}^n T_i,$7-step single-agent problems on a time-expanded graph and returns guidance paths $\mathrm{SoC}=\sum_{i=1}^n T_i,$8. For each agent $\mathrm{SoC}=\sum_{i=1}^n T_i,$9,

$T_i$0

with lexicographic edge costs

$T_i$1

PIBT then re-ranks candidate moves by

$T_i$2

strictly preferring the first hop of the guidance path. The paper states that this preserves completeness. On the maze-128-128-10 map with 1,000 agents, vanilla LaCAM reports flowtime $T_i$3 in $T_i$4, while LaCAM with local guidance reports $T_i$5 in $T_i$6. Across 644 instances and 32 maps, local guidance typically reduces flowtime by $T_i$7 over vanilla LaCAM at the cost of $T_i$8 of additional planning time, and in tests with up to 10,000 agents it maintains a $T_i$9 flowtime gain at runtimes below $i$0 (Arita et al., 21 Oct 2025).

The lifelong extension LLLG—Lifelong LaCAM with Local Guidance—embeds guidance-enhanced LaCAM in a receding-horizon loop with horizon shift $i$1. At each timestep it solves a finite-horizon MAPF instance, executes only the first joint move, assigns new goals to agents that have just reached theirs, and warm-starts the next guidance from the suffix of the previous plan. This warm-start mechanism is reported to outperform inheriting only the previous guidance. On a $i$2 warehouse map with 10,000 agents, LLLG with small windows $i$3 runs in under $i$4 per step and achieves an $i$5 throughput gain over plain PIBT; against RHCR, it is $i$6 faster while matching or outperforming throughput in dense maps (Arita et al., 16 May 2026).

5. Kinodynamic, warehouse, and control-oriented extensions

db-LaCAM extends LaCAM from discrete graph MAPF to multi-robot kinodynamic planning by combining motion primitives, discontinuity-bounded stitching, and a lightweight MAPF layer. Each robot $i$7 has continuous state $i$8, control $i$9, and discretized dynamics

$Q\in V^n$0

A finite library of motion primitives $Q\in V^n$1 is precomputed, and when the current state is $Q\in V^n$2, a primitive $Q\in V^n$3 is applicable if

$Q\in V^n$4

The planner searches over horizon-length primitive sequences; local coordination is handled by db-PIBT, and heuristic assignment uses HEST. The stated guarantee is probabilistic resolution-completeness: if primitive clustering samples each cluster with non-zero probability and $Q\in V^n$5, $Q\in V^n$6, and the primitive set are fixed, db-LaCAM will find a solution with probability tending to $Q\in V^n$7 as search time tends to $Q\in V^n$8, provided a solution exists at that primitive resolution. Empirically, db-LaCAM scales to scenarios with up to 50 robots, achieves up to ten times faster runtime than state-of-the-art planners while maintaining comparable solution quality, and is validated in 2D and 3D as well as in physical experiments with 10 quadrotors and 4 car-with-trailer robots (Moldagalieva et al., 7 Dec 2025).

A different adaptation targets warehouse AGVs through the multi-agent warehouse pathfinding (MAWPF) model. There, each agent state is $Q\in V^n$9 with straight-only motion, in-place rotations requiring $\sigma$0 steps per $\sigma$1, acceleration and deceleration constraints, and follower-collision prohibition over swept cells. The high-level LaCAM loop is unchanged, but the configuration generator is replaced by an $\sigma$2-step PIBT variant that enumerates feasible kinodynamic path fragments, reserves swept cells in a time-indexed table, and commits only the first step $\sigma$3 as the successor. The paper reports that committing all $\sigma$4 steps at once—the “macro-successor” style—almost always fails in MAWPF because of dead-end states in the directed kinodynamic state-graph. With $\sigma$5 and $\sigma$6, LaCAM+PIBT achieves $\sigma$7 success up to 200 agents on random and room maps, with runtime $\sigma$8 and normalized SoC $\sigma$9 lower-bound; on the width-1-corridor warehouse-20-40-10-2-1 map, success falls to $A=\{1,\dots,n\}$00, which the authors interpret as PIBT inheritance stalling in narrow passages (Nagai et al., 15 May 2026).

LaCAM has also been embedded as a discrete planner inside hierarchical control stacks. In LF, a centralized discrete full-horizon planner based on LaCAM runs asynchronously with robot-wise dynamics-aware trajectory controllers based on Freyja. The high level operates at $A=\{1,\dots,n\}$01 Hz, while the low level tracks continuous trajectories at $A=\{1,\dots,n\}$02 Hz. The framework is demonstrated with multirotors and ground robots, including deployment of 15 real multirotors under random, consecutive target updates while a person walks through the workspace, illustrating how LaCAM can function as the combinatorial coordination layer inside a larger control architecture (Shankar et al., 15 Jul 2025).

6. Empirical profile, limitations, and nomenclature

Across the MAPF literature, LaCAM’s empirical profile is defined by very high scalability, strong first-solution speed, and a persistent trade-off between completeness or eventual optimality and immediate solution quality. The original LaCAM paper reports that it solved all six small “hard” instances quickly, achieved near-$A=\{1,\dots,n\}$03 success up to dense $A=\{1,\dots,n\}$04 grids with 400 agents, and on the warehouse-20-40-10-2-2 benchmark was the only solver to solve all instances from 1,000 to 10,000 agents, all within $A=\{1,\dots,n\}$05 even at 10,000 agents. The LaCAM$A=\{1,\dots,n\}$06 paper, in turn, reports that LaCAM$A=\{1,\dots,n\}$07 sub-optimally solved $A=\{1,\dots,n\}$08 of the instances retrieved from the MAPF benchmark within ten seconds on a standard desktop PC while ensuring eventual convergence to optima (Okumura, 2022, Okumura, 2023).

The main limitations are also consistent across papers. In the original benchmarks, LaCAM’s SoC ratio is typically around $A=\{1,\dots,n\}$09, better than pure greedy methods such as PP and PIBT but worse than EECBS and LNS2 on solution quality. A handcrafted swap-in-corridor instance exposes a pathological regime in which high-level search over nearly identical configurations explodes: LaCAM takes 128 iterations for 2 agents, 23,907 for 4 agents, and 287,440 for 6 agents. In LaCAM$A=\{1,\dots,n\}$10, baseline initial solutions on 1,000-agent grids are reported to be often $A=\{1,\dots,n\}$11, and convergence after the first solution can be slow. Guidance methods improve quality but add measurable overhead; local guidance has complexity $A=\{1,\dots,n\}$12 per guidance call and typically adds a few seconds for $A=\{1,\dots,n\}$13 and $A=\{1,\dots,n\}$14. In AGV-style MAWPF, narrow passages remain difficult because PIBT-based inheritance can stall (Okumura, 2022, Okumura, 2023, Arita et al., 21 Oct 2025, Nagai et al., 15 May 2026).

Several recurring misconceptions are resolved by the variant structure of the literature. Base LaCAM is complete but sub-optimal; LaCAM$A=\{1,\dots,n\}$15 is eventually optimal only under cumulative transition costs; Real-Time LaCAM is real-time and provably complete because it preserves and continues to accumulate constraints across iterations, whereas naively rerunning LaCAM from scratch each step loses those constraints and risks livelock or deadlock. Similarly, not all “guidance” in the LaCAM ecosystem is static: local guidance is recomputed online around the current configuration, and the Lightweight Traffic Map is updated dynamically from PIBT’s own search history (Liang et al., 8 Apr 2025, Arita et al., 21 Oct 2025, Shen et al., 9 Mar 2026).

Finally, the acronym is not entirely unique. In MAPF and multi-robot planning, LaCAM denotes the lazy-constraint configuration-search framework described above. In an unrelated interpretability context, the name “LaCAM” is also used for a Layer-Adaptive CAM mechanism inside “Integrative CAM,” where it refers to adaptive multi-layer class activation mapping rather than multi-agent pathfinding (Singh et al., 2024).