Priority Inheritance with Backtracking for MAPF

• PIBT is a decentralized MAPF method that ensures collision-free joint moves using temporary priority inheritance and recursive backtracking.
• It computes one-step moves with bounded runtime, making it effective for iterative and lifelong planning in dynamic environments.
• Extensions enhance PIBT's performance through refined candidate ordering, artificial potential fields, and adaptations for non-biconnected graphs.

Priority Inheritance with Backtracking (PIBT) is a decentralized, rule-based method for multi-agent path finding (MAPF) that generates a collision-free next configuration by planning one timestep at a time in priority order, using temporary priority inheritance and local backtracking to resolve blocking chains. In the standard discrete-time graph model, agents synchronously choose either a move to an adjacent vertex or a wait action while avoiding vertex and swap conflicts. PIBT was introduced for iterative and lifelong settings in which bounded per-timestep runtime, local communication, and practical scalability are more important than global optimality, and it has since become a common low-level configuration generator in larger MAPF frameworks and post-processing pipelines (Okumura et al., 2019).

1. Definition and problem setting

PIBT operates on a graph $G=(V,E)$ with discrete synchronous timesteps, where each agent occupies one vertex and in one timestep may either remain in place or move to an adjacent vertex. A valid joint move must avoid vertex conflicts, in which two agents occupy the same vertex at the same time, and swap conflicts, in which two agents traverse the same edge in opposite directions in the same timestep (Okumura et al., 2019).

The algorithm is most naturally associated with iterative MAPF and lifelong MAPF rather than one-shot optimal planning. In the iterative setting, PIBT is called at every timestep to produce only the next joint action, after which the process repeats from the new configuration. In lifelong settings, agents repeatedly receive new destinations after reaching current goals, so the planner must sustain flow over long horizons rather than solve a single finite episode once and for all. This emphasis on repeated local replanning explains why PIBT is routinely described as a “configuration generator” or a “single-step solver,” rather than a full-horizon search method (Gandotra et al., 10 Apr 2025).

In the original graph-theoretic formulation, PIBT was designed for environments satisfying a cycle condition: all pairs of adjacent nodes belong to a simple cycle, with biconnected undirected graphs as a standard example. Under that condition, the method has a reachability guarantee suited to iterative operation: all agents eventually reach their destinations within finite time, even though the algorithm does not guarantee a classical one-shot MAPF solution in which all agents simultaneously remain at their goals (Okumura et al., 2019).

2. Core mechanism

At each timestep, PIBT maintains a strict priority order over agents and processes them from highest to lowest priority. Each agent forms a candidate set consisting of neighboring vertices plus the wait action, and those candidates are ordered by a local heuristic, typically shortest-path distance to the agent’s goal. The basic behavior is therefore greedy and goal-directed: absent interference, higher-priority agents take the locally best next step toward their goals (Okumura et al., 2019).

The distinctive part of PIBT is priority inheritance. If a high-priority agent wants to move into a vertex currently occupied by a lower-priority agent that has not yet fixed its next move, the lower-priority agent temporarily inherits the higher priority and is recursively forced to choose a new collision-free successor. This can induce a chain of inherited decisions in which one blocked agent pushes another, and so on. The inherited priority is temporary and local to the current planning step; it does not replace the base priority scheme across timesteps (Veerapaneni et al., 2024).

Backtracking resolves failures inside such chains. If an inherited agent cannot find any valid successor because all admissible moves either collide with already reserved moves or lead to recursive failure downstream, the recursion returns failure to its parent. The parent then abandons the current candidate and tries another one. If all candidates fail, the process backtracks further up the chain. In this way, PIBT performs a localized recursive search over the next joint configuration while still remaining a one-step method (Okumura et al., 2019).

This yields a characteristic division of labor. Priority ordering determines whose preferences dominate, priority inheritance propagates those preferences through blocking chains, and backtracking prevents the chain from committing to locally inconsistent assignments. The result is a collision-free joint action with very low per-timestep computational cost, but with strong dependence on priority order and local candidate ranking (Gandotra et al., 10 Apr 2025).

3. Guarantees, complexity, and known limitations

The original analysis gives PIBT a reachability guarantee on graphs where every adjacent pair of nodes lies on a simple cycle. Under that condition, every agent reaches its destination within finite time; the 2019 formulation states an upper bound of $\mathrm{diam}(G)\cdot |A|$ timesteps after destinations are given (Okumura et al., 2019). Later summaries describe the method as complete for iterative or pickup-and-delivery variants under suitable graph conditions, while still emphasizing that PIBT is not a complete solver for general one-shot MAPF (Veerapaneni et al., 2024).

A recurring theme in subsequent work is that PIBT is exceptionally fast because it plans only one step ahead. A later large-scale free-space study cites time complexity per timestep as

$O\bigl(|A|(\Delta(G)+\log |A|)\bigr),$

where $\Delta(G)$ is the maximum degree of the graph (Chakravarty et al., 20 Jun 2025). This computational profile is one reason PIBT is routinely used as a low-level component inside larger systems.

The same single-step design also creates its main weaknesses. PIBT is repeatedly described as extremely greedy with respect to both priorities and local heuristic ordering. It returns the first collision-free joint action it finds, cannot exploit extra planning time in its standard form, and may be arbitrarily suboptimal for the single-step objective used in later analyses (Gandotra et al., 10 Apr 2025). Because decisions are made only one timestep ahead, PIBT can behave shortsightedly in corridors and bottlenecks, and one-step local optimality does not reliably translate into improved full-horizon cost (Okumura et al., 2019).

Another limitation is structural. Original PIBT assumes an environment compatible with its cycle-based progress argument. In graphs with dead-ends or tree-shaped branches, deadlocks can arise. This motivated later extensions that preserve the PIBT decision pattern while altering priorities and permissible moves in non-biconnected regions (Fujitani et al., 2022).

4. Priority schemes and candidate ordering

Although PIBT’s mechanics are fixed, a large part of its empirical behavior is governed by how priorities and candidate preferences are constructed. The original formulation uses dynamic priorities so that agents that have not reached their goals for longer accumulate priority, while goal-reaching agents reset to a small tie-breaking value; this provides fairness and underlies the reachability proof in iterative settings (Okumura et al., 2019). Other papers summarize alternative practical formulations in which priority values are derived from distance-to-goal or related quantities, but the invariant is always a strict total order at each timestep (Yukhnevich et al., 12 Nov 2025).

Subsequent work shows that candidate ordering is particularly influential. A 2025 study on lightweight preference construction treats PIBT as a fixed local configuration generator whose only degree of freedom is the preference list used to sort each agent’s candidate actions (Okumura et al., 19 May 2025). In the vanilla form, the preference key is

$\langle \mathrm{dist}(v,g_i), \epsilon \rangle,$

so equal-distance actions are effectively broken randomly. That paper argues that many practical failures arise precisely because graph shortest paths are often non-unique, and random tie-breaking among equally good local moves has large downstream effects on congestion and solution cost (Okumura et al., 19 May 2025).

Two lightweight refinements are introduced there. “Hindrance” evaluates whether a candidate move is likely to block neighbors’ progress in the next timestep, while “regret” is learned across multiple PIBT runs and estimates how much a chosen action forces others away from their own best moves. These refinements do not alter priority inheritance or backtracking themselves; they only change candidate ranking. Empirically, the paper reports that in dense one-shot cases the combined tie-breaks improve sum-of-costs by around $10\%$–$20\%$, and in lifelong MAPF hindrance-based strategies can improve throughput by at least $40\%$ on some instances, without significantly compromising speed (Okumura et al., 19 May 2025).

This line of work clarifies an important point: much of PIBT’s practical quality is controlled not by the recursion structure alone, but by what local options are presented first to that recursion. Later extensions that incorporate guide paths, artificial potential fields, or learned action preferences all exploit this same leverage point.

5. Major extensions and reinterpretations

Several later methods preserve PIBT’s inheritance-and-backtracking core while changing its planning horizon, action model, or search interpretation.

winPIBT generalizes PIBT from a one-step window to a configurable time window. Agents reserve short path prefixes rather than only the next move, and priority inheritance can propagate across future timesteps. The method preserves finite-time reachability under analogous graph conditions and mitigates livelock situations seen in PIBT, especially in corridor-heavy maps, but the paper also shows that no single window size is uniformly best and that larger windows increase computation and communication range (Okumura et al., 2019).

Anytime PIBT reinterprets PIBT’s recursion as a depth-first search over the joint single-step action tree. It first returns exactly the same initial one-step solution as PIBT, then continues improving that solution in an anytime manner by exploring alternative joint actions and using disjoint agent groups for tractability. Given sufficient time, it converges to the optimal single-step solution, yet the reported experiments also show that better single-step solutions have only small effects on full-horizon costs (Gandotra et al., 10 Apr 2025).

CS-PIBT embeds PIBT as a collision shield for learned local policies. A neural policy proposes per-agent action probabilities, and PIBT-style one-step conflict resolution converts those proposals into a collision-free joint action. In this formulation, plain PIBT is the special case in which the “policy” simply prefers the greedy heuristic-minimizing action. The resulting studies argue that future learned MAPF policies should always be evaluated against PIBT-like shields because one-step collision resolution is already handled efficiently by CS-PIBT (Veerapaneni et al., 2024).

EPIBT extends PIBT to settings with orientations and time-consuming rotations by replacing single actions with short multi-action operations. It also introduces revisiting of agents within a timestep and operation inheritance across timesteps. The method preserves PIBT’s bounded-progress flavor under the same cycle condition while addressing cases where one-step planning is too myopic for rotation-based motion. Combined with graph guidance and large neighborhood search, it achieves state-of-the-art throughput in online lifelong MAPF with rotations in the reported experiments (Yukhnevich et al., 12 Nov 2025).

MD-PIBT provides a more abstract generalization. It introduces safe paths, tentative paths, hard dependencies, soft dependencies, and an Agent Dependency Graph, then treats PIBT as the special case in which each tentative path can conflict with at most one other agent, so the dependency graph degenerates to a chain. By allowing multiple dependencies per path, MD-PIBT can express planning behaviors that PIBT and EPIBT cannot, which is especially useful for large agents and more complex kinodynamic models (Jiang et al., 24 Mar 2026).

6. Applications and empirically observed regimes

A substantial body of later work uses PIBT not only as a standalone planner but as a module inside broader systems. In free-space path planning, an 8-connected version of PIBT combined with safety-aware string-pulling and local SIPP-based correction scales to over 500 agents, reduces path lengths while maintaining real-time performance, and produces trajectories reported as on average $0.8\%$ worse than CCBS+SIPP on solvable benchmark cases (Chakravarty et al., 20 Jun 2025). In that pipeline, PIBT supplies the discrete collision-free backbone, while geometric smoothing is restricted to regions where agents are not locally interacting.

In lifelong MAPF, congestion-aware modifications are particularly effective. A 2025 study integrates artificial potential fields into PIBT’s neighbor ranking by replacing the score $h(v)$ with

$\mathrm{diam}(G)\cdot |A|$0

where the additional term aggregates short-horizon repulsive potentials induced by already planned higher-priority agents. The paper reports that APFs are not beneficial for standard one-shot MAPF but can yield up to a 7-fold increase in overall system throughput for lifelong MAPF; on the room-32-32-4 grid, PIBT+APF achieves about $\mathrm{diam}(G)\cdot |A|$1 higher throughput than vanilla PIBT on average (Pertzovsky et al., 28 May 2025). The same paper notes that “In most cases, PIBT+APF outperforms other approaches, including L-PIBT+GP, and improves the original PIBT by a significant margin” in lifelong settings (Pertzovsky et al., 28 May 2025).

For non-biconnected environments, PIBT with temporary priority extends the algorithm to a biconnected main area plus attached tree-shaped dead-end regions. Agents inside a tree whose destination lies outside that tree are assigned temporary high priority, and movements within trees are restricted so that evacuation toward the main area remains possible. A further PIBTTP-TA variant introduces temporary avoidance on branches. Under the stated structural and task assumptions, both variants guarantee completion of all tasks in finite time, and the experimental results show PIBTTP-TA outperforming token passing across the tested MAPD environments (Fujitani et al., 2022).

Across these diverse extensions, the basic identity of PIBT remains stable. It is a fast local solver that commits to collision-free next moves through priority-ordered recursion, then relies on carefully chosen heuristics, short-horizon refinements, or post-processing layers to compensate for greediness and myopia. Its continuing importance in MAPF research lies less in optimality than in its role as a scalable primitive: a practical mechanism for transforming local preferences into feasible multi-agent motion at very high frequency (Gandotra et al., 10 Apr 2025).