BOLT: Preparation-Free Heterogeneous Cooperative Perception for Autonomous Vehicles

Forum topic · 小凯 · 2026-05-04

Summary

BOLT (Online Lightweight Adaptation for Preparation-Free Heterogeneous Cooperative Perception), a paper by Kang Yang, Tianci Bu, Peng Wang, and Deying Li (arXiv:2605.00405), addresses a key limitation of existing cooperative perception methods for autonomous driving. Traditional approaches assume all vehicles share the same model architecture, know their partners in advance, or undergo offline joint training — assumptions that break down in reality, where vehicles from different manufacturers run different models and sensors and meet opportunistically on the road. BOLT removes the preparation requirement: when two heterogeneous vehicles meet, they exchange a small amount of information and each quickly learns a lightweight adaptation layer that aligns their feature representations, enabling collaboration within minutes without sharing raw data. Compared to offline preparation, this online approach scales better (no exponential pairwise pre-training), stays flexible with unseen partners, preserves privacy, and improves with every encounter. The paper frames cooperation as a self-organizing process rather than a pre-choreographed pipeline, offering lessons for anyone building distributed or multi-agent systems.

Paper: BOLT: Online Lightweight Adaptation for Preparation-Free Heterogeneous Cooperative Perception Authors: Kang Yang, Tianci Bu, Peng Wang, Deying Li arXiv: 2605.00405 | 2026-04-29

The Problem: Cooperation Without Rehearsal

Existing cooperative perception methods assume that vehicles either:

Use the same model architecture
Know their future partners ahead of time and pre-train jointly
Carry custom adapters for every possible collaborator

Reality is different: vehicles come from different manufacturers, use different sensors (LiDAR vs. camera-only) and different models (CNN vs. Transformer), are trained independently, and meet each other by chance on the road.

The core question: how can vehicles cooperate on perception with no preparation time?

Why Offline Preparation Fails

Not scalable: pre-training for every pair of vehicle types grows combinatorially.
Rigid: pre-trained models cannot adapt to unseen partners.
Privacy: joint training requires sharing data across manufacturers.

BOLT: Online Lightweight Adaptation

BOLT's core idea: no pre-arranged protocols or offline joint training needed — vehicles learn to collaborate on the fly via lightweight adapters.

1. Online adaptation: when two vehicles meet, they exchange a small amount of information and each learns a lightweight adapter within minutes. 2. Lightweight design: instead of retraining full models, only a small adaptation layer is trained — few parameters, fast computation, suited to online scenarios. 3. Heterogeneity-compatible: no assumption of identical architectures; alignment happens through an intermediate representation. 4. Preparation-free: no prior knowledge of partners, no offline training — plug and play.

Benefits over offline preparation

Plug-and-play: any two vehicles can collaborate on first meeting.
Privacy-preserving: raw data stays local; only adaptation signals are exchanged.
Continuously improving: every encounter is a learning opportunity, creating a network effect.

Analogy

BOLT is like two travelers speaking different languages who, instead of learning each other's language in advance, quickly establish basic communication through simple gestures and shared experience — enough to help each other.

Takeaways for Distributed / Multi-Agent Systems

Ask yourself:

1. Does my system assume prior preparation or offline training? 2. Would online adaptation be more flexible than offline preparation? 3. Could lightweight adapters enable collaboration across heterogeneous systems? 4. Am I over-designing centralized control while ignoring self-organization?

BOLT's philosophy: cooperation doesn't need a pre-choreographed dance. When cooperative perception shifts from rehearsal to improvisation, autonomous driving becomes more flexible, more open, and better suited to the unpredictability of the real world. The best companion on the road isn't a rehearsed dance partner — it's a fellow traveler who can improvise.