CBS+SIPP in Multi-Agent Path Finding

• Conflict-Based Search with SIPP is a framework that integrates a high-level constraint-tree search with low-level safe interval planning to efficiently resolve multi-agent collision constraints.
• It employs safe interval representations and continuous-time adaptations to allow arbitrary travel durations and multi-objective optimizations while ensuring optimality.
• Enhanced techniques like disjoint splitting and heuristic conflict prioritization significantly improve scalability and runtime efficiency in complex multi-agent scenarios.

Conflict-Based Search (CBS) with Safe Interval Path Planning (SIPP) constitutes a principled and scalable methodology for optimal Multi-Agent Path Finding (MAPF) under both discrete and continuous-time, as well as multi-objective and non-unit cost regimes. By synthesizing CBS’s constraint-tree high-level search with SIPP-derived low-level planners that exploit safe-time-interval representations, these frameworks resolve complex agent-collision constraints with completeness and optimality guarantees. Recent extensions generalize CBS+SIPP to continuous-time domains (CCBS), edge- and interval-constrained scenarios, and multi-objective optimization settings, enabling application to a broad spectrum of real-world multi-robot systems.

1. Formal Basis: CBS and SIPP Integration

CBS operates as a two-level scheme for MAPF, with a high-level search over a constraint tree (CT) and per-agent low-level planners that respect the associated constraints (Andreychuk et al., 2019). In the classical, discrete-time CBS setting, constraints prohibit agents from occupying conflicted vertices or edges at integer timesteps; the low-level planner is typically A* on the time-augmented graph.

SIPP enhances per-agent planning by representing the accessibility of each graph node over a collection of safe intervals—maximal, collision-free time windows for occupancy or traversal. For continuous-time or multi-timestep settings, SIPP’s state-space expands to (vertex, interval) pairs, and safe intervals are dynamically updated at each CT node as global inter-agent constraints evolve (Andreychuk et al., 2019, Andreychuk et al., 2021). This mechanism not only alleviates the exponential state blowup of fine time discretization but permits arbitrary travel/wait durations and continuous collision detection.

2. Continuous-Time Conflict-Based Search (CCBS)

CCBS generalizes CBS+SIPP to continuous-time agent motion and geometric collision models. Each agent’s plan is a sequence of timed actions, either move or wait, each parametrized as (action, start time), and the joint solution cost (sum-of-costs, SOC) is $\sum_i \mathrm{cost}(\Pi_i)$ (Andreychuk et al., 2021).

Conflict detection in CCBS consists of identifying maximal unsafe intervals over which two agents’ actions would overlap in space. For each conflict $(a_i, t_i, a_j, t_j)$, maximal unsafe intervals $[t_i, t_i^u)$ and $[t_j, t_j^u)$ are computed for both agents. Resolving such a conflict spawns two child CT nodes by enforcing negative (forbid-action) constraints $\lnot(i, a_i, [t_i, t_i^u))$ and $\lnot(j, a_j, [t_j, t_j^u))$ respectively (Andreychuk et al., 2021, Andreychuk et al., 2019).

SIPP’s role in CCBS (often called GSIPP in this context) is to propagate these constraints into updated safe intervals and compute, via best-first search, the cost-optimal constraint-compliant single-agent path. The completeness and optimality of this reduction are guaranteed by soundness and minimality arguments on unsafe interval constraints (Andreychuk et al., 2019).

3. Enhancements: Disjoint Splitting, Conflict Prioritization, and Heuristics

The introduction of disjoint splitting, prioritized conflict resolution, and admissible heuristics significantly advances CCBS’s scalability while maintaining optimality (Andreychuk et al., 2021):

• Disjoint Splitting (DS): Instead of naively splitting on both negative constraints, DS selects one agent and introduces both a negative (“forbid $a_i$ on $[t_i, t_i^u)$”) and a dual positive constraint (“require $a_i$ on $[t_i, t_i^u)$”). This ensures the search subtrees are disjoint in solution space.
• Cardinality and Cost-Impact: Conflict prioritization is guided by a continuous cost-impact metric:

$(a_i, t_i, a_j, t_j)$0

where $(a_i, t_i, a_j, t_j)$1 denotes SOC at node $(a_i, t_i, a_j, t_j)$2 and $(a_i, t_i, a_j, t_j)$3 are child CT nodes. Conflicts are classified as cardinal (cost increases in both branches), semi-cardinal (cost increases in one branch), or non-cardinal (otherwise), and those with the largest cost impact are expanded first.

• High-Level Heuristics: Two admissible heuristics have been introduced:
  • $(a_i, t_i, a_j, t_j)$4: Minimum-vertex-cover linear program over agents and observed conflicts, ensuring that at least one agent per conflict is penalized by the minimal cost-impact—this gives an admissible lower bound.
  • $(a_i, t_i, a_j, t_j)$5: Greedy 2-approximate cover based on maximal uncoupled cost-impact conflicts.

These extensions reduce open list expansions, accelerate convergence, and demonstrably improve agent scalability (Andreychuk et al., 2021).

4. Generalizations: Multi-Objective and Non-Unit-Cost MAPF

Recent frameworks further extend CBS+SIPP concepts:

• Non-Unit Integer Costs: CBS-NIC adapts both the CT search and the SIPP subroutine to graphs $(a_i, t_i, a_j, t_j)$6 with $(a_i, t_i, a_j, t_j)$7 and maintains per-edge, per-agent time interval constraints. SIPP at the low level propagates constraint-derived safe intervals, and edge conflicts are handled by intersection checks over move-intervals. Empirical results demonstrate orders-of-magnitude improvements in runtime and success rate compared to both vanilla CBS and CCBS for large numbers of agents (up to 100+) (Fan et al., 7 Apr 2026).
• Multi-Objective MAPF: Multi-Objective Safe Interval Path Planning (MO-SIPP) pairs SIPP states with vector-valued costs, maintains per-state Pareto frontiers, and propagates vector dominance for multi-objective optimality. MO-CBS incorporates this MO-SIPP as its low-level planner, using Pareto-optimal joint paths at each CT node and enforcing conflict-derived constraints as in scalar CBS. MO-SIPP is proven to find all cost-unique Pareto-optimal paths per agent, and pruning is performed by vector dominance at both low and high levels (Ren et al., 2021).

5. Empirical Performance and Application Domains

CCBS and its variants have been benchmarked extensively on OMPL-generated roadmaps and canonical grids (including $(a_i, t_i, a_j, t_j)$8, warehouse, and den520d) (Andreychuk et al., 2021, Fan et al., 7 Apr 2026). Integrating DS, PC, and admissible heuristics nearly doubles the number of solved instances under fixed time budgets and achieves $(a_i, t_i, a_j, t_j)$9 success at 20 agents versus $[t_i, t_i^u)$0 for vanilla CCBS on dense roadmaps. Open list expansions are reduced by up to $[t_i, t_i^u)$1, and the best variants exceed vanilla CCBS by up to $[t_i, t_i^u)$2 in runtime efficiency.

Experiments on integer-cost graphs show that enhanced CBS-NIC+SIPP scales to 100+ agents at sub-second runtimes, while CCBS and discretized CBS fail beyond 20–30 agents (Fan et al., 7 Apr 2026).

6. Theoretical Guarantees and Correctness

All CBS+SIPP frameworks preserve:

• Completeness: Every conflict-free solution generated is globally feasible, owing to the exhaustiveness of constraint enumeration and the soundness of unsafe interval constraints (Andreychuk et al., 2019).
• Optimality: The lowest-cost, constraint-satisfying joint plan is found first. For multi-objective settings, the algorithms identify the entire Pareto front of non-dominated vectors (Ren et al., 2021).
• Disjoint Subtree Property: Disjoint splitting and positive/negative constraint duality guarantee non-overlapping solution spaces for search subtrees, preserving both optimality and pruning correctness (Andreychuk et al., 2021).

7. Landscape and Ongoing Developments

Conflict-Based Search with SIPP has established itself as the leading paradigm for solving MAPF in continuous time, arbitrary cost, and multi-objective domains. Ongoing research addresses fine-grained conflict detection, adaptive graph discretization (e.g., via Bayesian optimization), and further generalizations to richer agent dynamics and geometric constraints (Fan et al., 7 Apr 2026). The robustness of CBS+SIPP frameworks to variant objectives and domain requirements cements their relevance across multi-robot coordination, automated warehouse logistics, and dynamic resource allocation problems.