cs.ROSep 26, 2026
SaveCollisionGAT: Controller-Agnostic One-Step Collision Screening for Multi-Agent Motion
Organizations: McGill University Montreal, Canada
Abstract
Before a team of robots moves, each proposed step must be checked for collisions with other robots and with obstacles. We present CollisionGAT, a graph-attention network that reads the current and proposed states of moving agents together with locally relevant stationary obstacles and returns one collision-risk score per moving agent. Any controller can use these scores to accept, repair, replan, or postpone a proposed step. We mount CollisionGAT on a continuous path-following controller and on GATeD, an obstacle-blind D* Lite planner that uses typed vetoes to update its planning graphs. Exact geometric checks supply the training labels and independently audit every executed step.
Figures & tables
Fig. 1: Agent–agent collision cases in continuous and grid motion. Dashed circles indicate current positions and solid discs or cells indicate proposed positions. The continuous swept label detects (a) endpoint overlap (subcase of a crossing overlap) and (b) crossing between safe endpoints. Under four-connected grid motion, (c) duplicate destinations and (d) head-on swaps are the only possible one-step agent collisions; every participant is labeled.
Fig. 2: CollisionGAT attaches between any controller and execution: safe joint proposals pass through, and a veto sends the controller back to repair and re-propose. The controller itself is unchanged.
Fig. 3: CollisionGAT in one picture. (a) Black arrows are proposed actions; blue arcs are receiver ’s attention over its senders (Eq. 10 ), with thickness proportional to weight. Agent proposes ’s destination and is emphasized by the illustrative attention weights. (b) The four-head summary is combined with the agent’s own state and mapped to one risk score (Eq. 11 ).
Fig. 4: Full trajectories of two CollisionGAT mounts. Left: independently planned continuous waypoint routes create an agent–agent crossing conflict; CollisionGAT vetoes the proposal before controller repair. Right: obstacle-blind GATeD uses obstacle-attributed vetoes to discover and share wall cells before repairing D ∗ Lite paths. Squares and stars denote starts and goals.
| Setting | # Worlds | Agents arrived | Collisions |
|---|---|---|---|
| Discrete GATeD | 100 | 995/995 | 0 |
| Continuous CollisionGAT-Guard | 100 | 846/994 | 0 |
TABLE I: Controller outcomes on the two 100-world test suites.