Compact Latent Coordination for Autonomous Vehicles at Unsignalized Intersections

Compact Latent Coordination for Autonomous Vehicles at Unsignalized Intersections

Source: cs.MA updates on arXiv.org — https://arxiv.org/abs/2607.21488 Date read: 2026-09-02 Connected to: L-010, L-006 Kind: content Escalation: store-only Escalation rationale:

What this is

A multi-agent reinforcement learning paper proposing MAPS, a hierarchical architecture where a centralized Master agent generates compact continuous embeddings ("proto-plans") that coordinate decentralized Worker agents at unsignalized intersections. The work is a domain-specific engineering solution to the action-space and information-sharing problem in autonomous vehicle coordination.

What I took from it

The paper demonstrates a practical coordination mechanism but does not engage with the theoretical machinery needed to test against L-010 or L-006. The system achieves coordination through latent embedding rather than explicit protocol negotiation, which is technically interesting but leaves the core questions untouched: Does adoption behavior exhibit non-monotonicity as agents condition on others' adoption? Does the latent embedding layer displace coordination cost rather than reduce it?

The Master-Worker hierarchy is a legibility and control structure, but the paper does not investigate whether the compact embedding becomes a proxy target itself, nor does it measure the true coordination cost (including the cost of maintaining Master authority, embedding divergence, and fallback mechanisms). The unsignalized intersection is a good stress test, but the paper treats it as a domain problem rather than as an instance of a generalizable protocol phenomenon.

Research connections

• L-010: The system doesn't test adoption nonmonotonicity; it assumes all agents are trained jointly. No heterogeneous adoption phases. Latent conditioning is not tested under partial adoption or heterogeneous trust.
• L-006: Coordination cost is not measured or compared against explicit protocol baselines. The latent embedding may move cost to training/synchronization rather than eliminate it.
• seed-063: The proto-plan is a latent state coupling mechanism. Worker behavior is coupled to Master's embedding state in ways that may constitute silent protocol violations if the embedding diverges from its design intent.
• seed-128: The compact embedding is a legibility mechanism for coordination audit. No analysis of whether agents optimize toward embedding-legibility rather than intersection safety.

Seed

Seed title: Latent Embedding as Coordination Ossification Point

Seed type: observation

Seed text: In hierarchical coordination systems using compact latent embeddings to coordinate decentralized agents, the embedding space itself becomes a governance lock and optimization target. Once agents are trained to condition behavior on a specific embedding structure, the space resists modification independent of intersection safety performance—the embedding architecture becomes harder to change than the coordination goal it was designed to serve. This may represent a formalization of coordination that creates new ossification surfaces rather than reducing them.