---

📰 News

• 2026-05   Paper submitted to a peer-reviewed venue.
• 2026-05   Code is publicly released 🎉
• 2026-05   arXiv preprint will be linked here once the ID is assigned.

---

✨ TL;DR

Most "heterogeneous cooperative perception" methods assume that all agents pre-train together — shared protocols, joint optimization, or collaborator-specific calibration data. This breaks the moment two agents from different vendors meet on the road.

BOLT assumes nothing of the sort. Each agent ships with its own independently trained single-agent detector. When two agents meet, BOLT inserts a tiny adaptive plugin between the neighbor's feature stream and the ego's frozen fusion module, and adapts the plugin online via ego-as-teacher distillation. No labels. No cooperative training data. No shared protocol.

Across DAIR-V2X and OPV2V, with multiple LiDAR/camera encoder pairs and multiple fusion strategies, BOLT consistently turns degraded preparation-free cooperation into useful cooperation that surpasses ego-only performance.

---

✅ Highlights

	Feature	Description
1	🆓 Preparation-free	Each agent uses its own independently trained detector. No joint training. No shared protocol.
2	🔄 Online adaptation	Plugin parameters are updated at deployment by single-pass test-time distillation.
3	🪶 Lightweight	About 0.9M trainable parameters. Encoders, fusion module, and detection head are all frozen.
4	🏷️ Label-free, data-free	Ego predictions act as teacher — no labels, no cooperative training data.
5	🧩 Plug-and-play	Works across PointPillars / SECOND / Lift-Splat-Shoot encoders and across multiple fusion strategies.

---

📊 Results

Evaluated on DAIR-V2X (real V2I) and OPV2V (simulated V2V).

In the preparation-free setting, vanilla unadapted fusion typically falls below ego-only detection — cooperation actively hurts. BOLT reverses this:

• 🚀 Up to +32.3 AP@50 over unadapted fusion in the preparation-free setting.
• 📈 Surpasses ego-only on every evaluated encoder pair across DAIR-V2X and OPV2V.
• 🪶 Trains only ~0.9M parameters per plugin.

  

_{Left: performance across encoder pairs — vanilla fusion (red) drops below ego-only, BOLT (blue) consistently surpasses it.    Right: AP@50 improves online as more frames stream in.}

Qualitative BEV — with vs. without BOLT

_{Each pair shares the same scene; the left tile uses vanilla cross-agent fusion, the right tile uses BOLT. Green = ground truth, red = predictions. Without BOLT, fused predictions drift, duplicate, or miss objects in the neighbor's view; with BOLT, predictions snap back onto the true objects.}

See the paper for full tables, ablations, and additional qualitative results.

---

⚙️ Installation

git clone https://github.com/sidiangongyuan/BOLT.git
cd BOLT

# Conda env
conda create -n bolt python=3.8 pytorch==1.12.0 torchvision==0.13.0 cudatoolkit=11.6 \
  -c pytorch -c conda-forge
conda activate bolt

# Python packages
pip install -r requirements.txt
pip install spconv-cu116                 # match your CUDA version

# Project-local extensions
python setup.py develop
python opencood/utils/setup.py build_ext --inplace

The codebase reuses some infrastructure from HEAL and OpenCOOD; if you run into platform-specific issues with spconv / cumm, those repos' troubleshooting tips also apply here.

---

📦 Data Preparation

We evaluate on the following datasets:

Dataset	Used	Source
DAIR-V2X-C	✅	DAIR-V2X with complemented annotations
OPV2V	✅	OpenCOOD (also additional-001.zip for camera)
OPV2V-H	⏳ TODO	Hugging Face — not used in this work; planned for future support.
V2X-Real	⏳ TODO	Official site — not used in this work; planned for future support.

Each dataset's directory layout is configured through the YAML files under opencood/hypes_yaml/. Update the dataset paths there before training or inference.

---

🚀 Quick Start

BOLT's deployment-time contribution is a single command: a frozen detector pair plus an online-adapted plugin. Before running it, you need a trained base model (single-agent detectors and a fused backbone). The minimal reproducible flow is:

Step 1 — Train your single-agent detectors

Train one detector per modality / encoder using your own data. Configs are under opencood/hypes_yaml/:

python -m opencood.tools.train \
  -y opencood/hypes_yaml/dairv2x/HEAL/lidar_pyramid_local.yaml

Repeat for each agent's modality (e.g., LiDAR PointPillars, LiDAR SECOND, camera Lift-Splat-Shoot). These detectors are trained independently — no cross-agent coordination.

Step 2 — Build the heterogeneous base checkpoint

Assemble the single-agent encoders into a base model that BOLT will plug into. Encoders, the fusion module, and the detection head will be frozen afterwards.

python -m opencood.tools.train \
  -y opencood/hypes_yaml/dairv2x/HEAL/lidar_pp_second_stage2.yaml \
  --stage1_model_dir <path_to_stage1_checkpoints>

Step 3 — Run BOLT online adaptation 🟢 (this is BOLT)

At deployment, the ~0.9M-parameter plugin is the only trainable component. It is updated online: one gradient step per incoming test sample, supervised by the ego detector's high-confidence predictions (no labels, no cooperative training data).

python -m opencood.tools.online_adapt \
  --model_dir <path_to_base_checkpoint> \
  --output_dir <output_path> \
  --lr 1e-4 --epochs 1 \
  --teacher_conf_thresh 0.3 \
  --boost_weight 0.1 --boost_lo 0.1 --boost_hi 0.3

Evaluate

python -m opencood.tools.inference --model_dir <path_to_checkpoint>

For ready-made runners, see scripts/inference/inference.sh and the multi-agent demos under scripts/more_agents/.

---

🗂 Project Structure

opencood/
├── models/
│   ├── plugin/                       # BOLT adaptive plugin (AdaIN + residual + gate)
│   ├── heter_pyramid_collab.py       # Heterogeneous pyramid model
│   ├── heter_encoders.py             # Multi-modality encoder registry
│   └── fuse_modules/                 # Fusion strategies (pyramid, attention, ...)
├── tools/
│   ├── online_adapt.py               # 🟢 Online TTT with ego-as-teacher distillation
│   ├── train.py                      # Standard training
│   └── inference.py                  # Evaluation
├── hypes_yaml/                       # Configs for DAIR-V2X, OPV2V
└── data_utils/                       # Dataset loaders + pre/post processors

scripts/
├── inference/                        # Off-the-shelf inference
├── train/                            # End-to-end training
└── more_agents/                      # Multi-agent (3+ cars) assembly + adaptation

## 🙏 Acknowledgements

This codebase is built on top of [**HEAL**](https://github.com/yifanlu0227/HEAL) and [**OpenCOOD**](https://github.com/DerrickXuNu/OpenCOOD). We thank the authors of DAIR-V2X, OPV2V, OPV2V-H, and V2X-Real for releasing the datasets that made this work possible.

---

📖 Citation

If you find BOLT useful, please cite:

@article{bolt2026,
  title   = {BOLT: Online Lightweight Adaptation for Preparation-Free Heterogeneous Cooperative Perception},
  author  = {Yang, Kang and Bu, Tianci and Wang, Peng and Li, Deying and Wang, Yongcai},
  journal = {arXiv preprint 2605.00405},
  year    = {2026}
}

📄 License

Released under the MIT License.