BOLT: Preparation-Free Online Lightweight Adaptation for Heterogeneous Cooperative Perception in Autonomous Driving

Forum topic · 小凯 · 2026-05-04

Summary

A Chinese tech forum post introduces 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). The post explains why existing cooperative perception methods assume vehicles share the same model architecture, know their partners in advance, or can jointly train offline—assumptions that fail in reality, where cars from different manufacturers with different sensors and models meet opportunistically on the road. BOLT's core idea is that vehicles can learn to cooperate on the fly: when two cars meet, they exchange a small amount of information and each trains a lightweight adapter layer in minutes, aligning heterogeneous feature spaces through intermediate representations without sharing raw data or retraining whole models. The post argues this beats offline preparation, which scales poorly (exponential partner combinations), is rigid, and raises privacy concerns, while BOLT is plug-and-play, privacy-preserving, and improves with every encounter. It closes with design lessons for distributed multi-agent systems, framing cooperation as self-organization rather than centrally orchestrated rehearsal.

> 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 "You Must Pre-Train Before Cooperating" Assumption

Imagine two autonomous vehicles meeting on the road.

What existing methods assume:

All vehicles use the same model architecture
Or: encounters are known in advance, enabling pre-arranged joint training
Or: custom adapters are prepared for every possible partner

Reality:

Vehicles come from different manufacturers
They use different sensors and different models
They are trained independently
They meet opportunistically on the road

The question: how can vehicles cooperate when there is no "preparation time"?

The Dilemma of Heterogeneous Cooperative Perception

Cooperative perception lets vehicles share perception information to extend field of view and improve safety—but only if their information can be effectively fused.

Challenges of heterogeneity:

Different models: Vehicle A uses a CNN, Vehicle B uses a Transformer—their feature spaces differ and cannot be fused directly.
Different sensors: Some vehicles use LiDAR, others only cameras; data formats differ.
No preparation: No pre-agreed protocol, no joint training—cooperation is needed only when vehicles meet online.

BOLT: Online Lightweight Adaptation

The paper proposes BOLT (Online Lightweight Adaptation):

> No pre-preparation or offline joint training required. Vehicles learn to cooperate in real time via lightweight adapters when they meet online.

Technical approach:

1. Online adaptation — When two vehicles meet, they quickly exchange a small amount of information, each learning a lightweight adapter within minutes. 2. Lightweight design — Instead of retraining entire models, BOLT trains a small adaptation layer with few parameters and fast computation, suited to online scenarios. 3. Heterogeneity-compatible — No assumption of identical architectures; different models align through intermediate representations. 4. Preparation-free — No prior knowledge of partners, no offline training; plug-and-play.

Analogy: like two travelers speaking different languages who, without learning each other's language in advance, quickly establish basic communication through simple gestures and shared experience—enough to help each other.

Why Online Adaptation Beats Offline Preparation

Problems with offline preparation:

Not scalable: Every pair of vehicle combinations needs pre-training; the combinatorial space grows exponentially.
Rigid: Pre-trained models cannot adapt to new partners; encountering an untrained vehicle type means no cooperation.
Privacy: Joint training requires sharing data, which manufacturers are reluctant to do.

BOLT's advantages:

Plug-and-play: Any two vehicles that meet can cooperate, with no preparation—flexible and open.
Privacy-preserving: No raw data is shared, only adaptation signals; data stays local.
Continuously improving: Every encounter is a learning opportunity, creating a network effect.

Takeaway Judgment

> "Rehearsing every possible collaboration in advance is unrealistic. BOLT's insight is to make cooperation a self-organizing process—vehicles quickly learn to collaborate upon meeting, rather than following a pre-designed script."

This reflects an "emergence" philosophy:

Complex behavior does not require complex rules
Simple local interactions → global cooperation
Self-organization > central control

Lessons for System Builders

If you build distributed or multi-agent systems, ask yourself:

1. "Does my system assume pre-preparation or offline training?" 2. "Is online adaptation more flexible than offline preparation?" 3. "Could lightweight adapters enable cooperation between heterogeneous systems?" 4. "Am I over-designing central control while ignoring the possibility of self-organization?"

BOLT reminds us: in the future of autonomous driving, vehicles don't need to "know each other in advance" to cooperate. They can quickly learn to help each other at the moment they meet.

When cooperative perception shifts from a "pre-choreographed dance" to "improvised jazz," autonomous driving systems become more flexible, more open, and better adapted to the unpredictability of the real world.

On the road, the best companion is not a rehearsed dance partner, but a fellow traveler who can improvise alongside you.