Decentralized Model Predictive Control of Connected and Automated Vehicles with Coupled Safety Constraints

arXiv:2607.11403 2026 Dynamics 1 ideas extracted · analyzed Aug 30, 2026

What the math gives to ML

The paper offers a constructive way to handle coupled, nonconvex pairwise safety constraints in a decentralized controller: replace each pairwise collision constraint by a locally enforceable half-space derived from a buffered Voronoi partition around announced neighbor trajectories. Its second transferable mechanism is iterative non-cooperative coordination, with Jacobi updates using stale peer plans and Gauss-Seidel updates using the newest available plans. A strong neural-network transfer is a safety-filtered decentralized multi-agent policy or world model in which each agent proposes an action, then projects it onto buffered Voronoi constraints induced by peer predictions. The key falsifiable signature is that sufficiently large safety buffers eliminate violations while the decentralized fixed-point iteration becomes stable only when the effective cross-agent coupling has spectral radius below one.

Ideas from this paper

Unverified 2026

Buffered Voronoi Safety Projection

Add a decentralized safety layer to a multi-agent neural policy or learned world model. Each agent first predicts an action or short trajectory, then projects its proposal into a half-space defined by each neighbor's announced trajectory and a positive buffer, avoiding a centralized nonconvex collision solve. Use Jacobi or Gauss-Seidel iterations when agents mutually revise their predicted trajectories.

Useful6/10
Difficulty5/10
Novelty6/10
Paper: Decentralized Model Predictive Control of Connected and Automated Vehicles with Coupled Safety Constraints arXiv:2607.11403