GD-RHCR framework scales multi-agent pathfinding to 10x more agents
New parallel planning method proves near-optimal with 90% lower compute cost
What if lifelong multi-agent pathfinding could keep its near-optimal quality while scaling to far more agents? A new framework does exactly that. The paper first proves that Rolling-Horizon Collision Resolution (RHCR) is near-optimal in a discounted MDP formulation. It then extends it into Group Decentralized RHCR (GD-RHCR), which splits agents into groups using a transitive communication scheme and plans each group in parallel. Both methods achieve exponentially close to optimal guarantees, revealing a duality between RHCR's time-based restrictions and GD-RHCR's space-based partitioning. Across varied maps, GD-RHCR delivers high throughput at higher agent counts with significantly lower per-plan cost.
- Proves RHCR is near-optimal in discounted MDP formulation of L-MAPF
- GD-RHCR partitions agents by transitive communication for parallel planning
- Achieves similar optimality guarantees as RHCR but scales to higher agent counts with lower per-plan cost
Why It Matters
Enables warehouse robotics and autonomous fleets to scale coordination without exponential compute overhead.