Research map / Multi-agent systems and games / Human-AI and goal-conditioned
CoPEFT: Fast Adaptation Framework for Multi-Agent Collaborative Perception with Parameter-Efficient Fine-Tuning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CoPEFT, a parameter-efficient fine-tuning framework (Collaboration Adapter plus Agent Prompt) that adapts pretrained multi-agent collaborative 3D detectors to new deployment data by updating under 1% of parameters.
Problem
Collaborative perception models trained on one data distribution degrade on new deployment environments, and existing adaptation methods (domain adaptation, full fine-tuning) are too costly for resource-constrained agents.
Approach
A convolutional Collaboration Adapter with a collaboration-derived modulation score adds specific knowledge to frozen BEV features (macro level), and an Agent Prompt initialized from the adapter output acts as a virtual agent in fusion (micro level), with only adapters, prompt and decoder trained.
Key results
- On OPV2V-to-DAIR-V2X adaptation with CoAlign, CoPEFT reaches AP@70 0.418 versus 0.217 for no adaptation and 0.340 for DUSA, using about 111K trainable parameters out of 13M.
- Averaged over settings, CoPEFT improves detection performance by 19.1% over the unadapted baseline at 10% labeled data.
- Ablation: Collaboration Adapter alone gives AP@70 0.408 and Agent Prompt alone 0.372, combined 0.418 versus 0.217 without adaptation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add CoPEFT adaptation | CoAlign · DAIR-V2X (OPV2V → DAIR-V2X) | AP@70: 0.217 -> 0.418 | Helped |
| add CoPEFT adaptation | AttFuse · DAIR-V2X | AP@70: 0.203 -> 0.374 | Helped |
| add CoPEFT adaptation | MKD-Cooper · DAIR-V2X | AP@70: 0.157 -> 0.362 | Helped |
| add CoPEFT adaptation | CoAlign · V2XSet (OPV2V → V2XSet) | AP@70: 0.839 -> 0.854 | Helped |
| add Collaboration Adapter and Agent Prompt (ablation) | CoAlign · DAIR-V2X (component ablation) | AP@70: 0.217 -> 0.418 | Helped |
| add Agent Prompt to the Collaboration Adapter (ablation) | CoPEFT · DAIR-V2X (component ablation) | AP@70: 0.408 -> 0.418 | Helped |
| add Collaboration Adapter to the Agent Prompt (ablation) | CoPEFT · DAIR-V2X (component ablation) | AP@70: 0.372 -> 0.418 | Helped |
| add convolution, Collaborative Filter, and Score Generator to the adapter (ablation) | Collaboration Adapter · DAIR-V2X (adapter ablation) | AP@70: 0.380 -> 0.408 | Helped |
| add convolution to the adapter (ablation) | Collaboration Adapter · DAIR-V2X (adapter ablation) | AP@70: 0.380 -> 0.389 | Helped |
| add Collaborative Filter to the adapter (ablation) | Collaboration Adapter · DAIR-V2X (adapter ablation) | AP@70: 0.380 -> 0.392 | Helped |
| add Score Generator to the adapter (ablation) | Collaboration Adapter · DAIR-V2X (adapter ablation) | AP@70: 0.380 -> 0.395 | Helped |
| add instance awareness to the prompt (ablation) | Agent Prompt · DAIR-V2X (prompt ablation) | AP@70: 0.308 -> 0.360 | Helped |
Surprising findings
- Training from scratch on limited deployment data performs worse than no adaptation at low availability rates (e.g., AP@70 0.053 at 1%).
- Standard linear adapters perform poorly at 1% data (AP@70 0.089), below the unadapted baseline (0.217).
Limitations
- Relies on a labeled deployment subset, requiring annotation effort even if small.
- CoPEFT_S variant with only intermediate adaptation yields no performance advantage; the fuller CoPEFT_D variant doubles parameters.
Experimental setup
Datasets: OPV2V DAIR-V2X V2XSet
Benchmarks: DAIR-V2X V2XSet
Metrics: AP@50 AP@70
Baselines: Training from Scratch Decoder Network only SSF Adapter MACP DUSA
Builds on: CoAlign Adapter SSF MACP DUSA
Closest papers by meaning
- An Extensible Framework for Open Heterogeneous Collaborative Perception (ICLR 2024 · similarity 0.91)
- NegoCollab: A Common Representation Negotiation Approach for Heterogeneous Collaborative Perception (NeurIPS 2025 · similarity 0.89)
- Pragmatic Heterogeneous Collaborative Perception via Generative Communication Mechanism (NeurIPS 2025 · similarity 0.88)
- Communication-Efficient Collaborative Perception via Information Filling with Codebook (CVPR 2024 · similarity 0.87)
- CoSDH: Communication-Efficient Collaborative Perception via Supply-Demand Awareness and Intermediate-Late Hybridization (CVPR 2025 · similarity 0.87)
- One is Plenty: A Polymorphic Feature Interpreter for Immutable Heterogeneous Collaborative Perception (CVPR 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.