At a Glance
A lightweight predictor can tell when replanning a fleet of robots will actually save time — recovering 94.6% of the available delay savings while avoiding needless replans most of the time.
ON THIS PAGE
What They Found
Randomly replanning as delays occur rarely helps and can sometimes hurt. A regression model using 18 execution-aware features (recent action delays, propagated waiting time called "slack", etc.) flags the uncommon but high-impact situations where replanning pays off. On a 12,000-run simulated benchmark with a single unobserved obstacle, the model recovered most of the possible savings with only one replanning opportunity and kept false-positive costs negligible.
Data Highlights
1Introducing a random blocking obstacle raised average total execution time by 5.6% (from 347.99s to 367.74s).
27.28% of runs had large recoverable slowdowns (up to about 200 seconds) where replanning could help significantly.
3The learned predictor recovered 94.6% of the available execution-time savings across the test set with a single replanning decision.
What This Means
Engineers running fleets of warehouse robots or delivery drones — use the signal to avoid spinning costly replans while still catching rare big disruptions. Technical leads and SREs for multi-agent systems — reduce wasted compute and operator overhead by triggering replans only when they matter. Researchers building robust planners — the feature set pinpoints where execution delays propagate and which signals matter most.
Not sure where to start?Get personalized recommendations
Key Figures
Fig 1: Figure 1 : Example MAPF plan. Full circles are agents, hatched circles their goals, and lines represent paths.
Fig 2: (a) MAPF solution
Fig 3: (a) random-32-32-20
Fig 4: Figure 4 : SOC across all scenarios
Ready to evaluate your AI agents?
Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.
Learn MoreYes, But...
Results come from simulation on moderate-size maps with a single unobservable dynamic obstacle; performance may change on much larger maps or with multiple/partially observable disturbances. The study evaluates a single replanning decision per run — sequential or repeated replanning strategies were not explored. Planning time was small in these experiments; on larger environments the computational cost of replanning could change the trade-offs observed here. replanning decision
Methodology & More
A practical signal can decide whether to replan a fleet of agents during execution. The approach monitors a running execution controlled by an action-dependency graph (a structure that enforces safe action ordering) and extracts 18 features that describe current delays and how those delays are likely to propagate (for example, the increase in "slack", meaning extra waiting time downstream). The authors generated 12,000 simulated runs where a single unobserved obstacle randomly blocked a vertex, collected the execution traces, and labeled each observation with the measurable time savings that would result from replanning at that moment.
A simple regression model trained on those features predicts the achievable time savings if the system replans now. Randomly triggering replanning showed almost no average benefit, but a small fraction of runs incurred very large delays that could be recovered by replanning. The predictor reliably identified those high-impact cases: with a single replanning opportunity it recovered 94.6% of available savings while causing negligible losses from false positives. Feature analysis shows recent action delays and increases in propagated waiting time drive decisions; static map or plan attributes matter much less. The method is planner-agnostic and remains effective when including the modest replanning overhead measured in these experiments. Future work should test sequential replans, larger maps where planning is costlier, and richer disturbance models (multiple or partially observable obstacles). action-dependency graph and replanning are central concepts in this analysis.
Avoid common pitfallsLearn what failures to watch for
Credibility Assessment:
Authors are from a recognized research university (Czech Technical University / CIIRC) but have low h-indices and the paper is an arXiv preprint rather than a top conference/journal; overall a solid but not top-tier signal.