Competition research · July 2026
First place in the League of Robot Runners 2026 combined evaluation.
The AIM Labs team — Egor Yukhnevich, Yerasyl Kenesbek and Yersain Kenesbek — competed as No Man’s Sky. Their final system led the combined evaluation with a score of 10.628 and completed 310,466 tasks.
- Result
- 1st
- Tasks completed
- 310,466
- Combined score
- 10.628
- Submissions
- 87
The challenge
Planning while the world keeps moving.
The competition combines lifelong multi-agent path finding with online task allocation. Agents execute multi-tick actions under stochastic delays, sparse state exchange and strict per-tick compute limits. A useful plan must be safe, fast and still match the state in which it will actually run.
The system
One online planning pipeline.
The final system treats prediction, assignment, search and compute budgets as one operating loop. Planning overlaps execution without losing a simulator-faithful view of the agents and their tasks.
- World model
- Replays the simulator forward and tracks delays, actions, task progress and future planning boundaries.
- EPIBTX
- Builds feasible short operation sequences with priority inheritance, backtracking and multi-tick reservations.
- Adaptive search
- Parallel Large Neighborhood Search improves a safe incumbent while respecting a hard live deadline.
- Scheduling and guidance
- A persistent background scheduler and map-aware heuristics model assignment, orientation, distance and congestion.
The result
The highest combined score.
The final configuration was selected through 87 competition submissions. It completed 310,466 tasks and recorded the highest combined score in the evaluation: 10.628. The development trace showed that prediction fidelity, exact map-aware heuristics and explicit compute-budget control mattered more than generic tuning.
Paper
Scalable lifelong multi-agent planning.
The full paper documents the architecture, development trace, component comparisons and limitations of the approach.
Read the paperResult wording reflects the combined evaluation available at the time of publication.