Skip to main content
Back to timeline
arXivSource publication:

CoopDistill adds a learned correction layer over fixed-path coordination, lifting roundabout success 27.1 points above the potential-game baseline

Synopsis

CoopDistill splits cooperative driving into two layers: a potential game supplies a fixed-path longitudinal coordination reference, and a residual multi-agent reinforcement learning policy learns longitudinal acceleration, conflict-approach timing and lateral path corrections around it. Across merging, roundabout and intersection scenarios at four connected-vehicle penetration levels from 25% to 100%, success rises by 2.1 to 27.1 percentage points and delay falls by 7.3% to 27.3% relative to the potential-game baseline, with no observed collisions. At execution, a finite-horizon rollout compares the learned proposal with the reference and falls back whenever feasibility or utility conditions fail.

Source-provided article image: CoopDistill: Equilibrium-Guided Multi-Agent Policy Distillation for Cooperative Driving
arXiv

Interpretation

Cooperative driving no longer relies only on longitudinal coordination along fixed paths; connected vehicles can change local paths while keeping the reference's coordination, releasing vehicles that block each other. Earlier optimization and potential-game methods simplified multi-vehicle planning with fixed paths, which left spatial rearrangement outside the feasible action set. Across merging, roundabout and intersection scenarios at four connected-vehicle penetration levels from 25% to 100%, success rises by 2.1, 27.1 and 13.1 percentage points respectively over the potential-game baseline, with no observed CAV-involved collisions.

The learned component does not generate the whole joint maneuver from scratch; it adds three correction channels on top of the reference solution. The channels are a longitudinal acceleration residual, a conflict-approach timing residual and a lateral path residual; the first two modify the longitudinal request, and the third uses a smooth offset profile that takes a vehicle off its path and back, leaving the destination unchanged. During training, the potential game runs scenes alone, and scenes are ranked by collision, completion and delay; poorly resolved or substantially delayed scenes enter a difficult subset that is mixed with fresh scenes, with the difficult share rising from 0.25 to 0.50 over training.

Before execution, the system rolls out the learned proposal and the potential-game reference from the same state and accepts the proposal only if it passes the checks. The checks reject predicted CAV-involved collisions, invalid motion, any-vehicle collisions absent from the reference prediction, and proposals whose rollout utility falls below the reference within tolerance. On the formal test sets, learned proposals are accepted for 69.2% of decisions on average and the remaining 30.8% fall back to the potential game; averaged across scenarios, 14.3% of episodes turn from potential-game failure to success, while the reverse transition occurs in 1.0%.

Perspective

The results come from simulated merging, roundabout and intersection scenarios with 120 common cases per scenario and penetration level, six to twelve vehicles per scene, and human-driven vehicles using heterogeneous car-following and yielding behavior models; the authors propose extending the work to sustained arrivals, interacting bottlenecks, uncertain driver responses and rollout decisions under communication delays.

The authors list sustained traffic with continuous arrivals, interacting bottlenecks, uncertain driver responses and communication delays as future work, so the current results are limited to single interactions over a finite horizon; the selector's comparison is local to the current state and prediction horizon, and later decisions depend on paths already committed.