Research map / Multi-agent systems and games / Human-AI and goal-conditioned
NegoCollab: A Common Representation Negotiation Approach for Heterogeneous Collaborative Perception
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 NegoCollab, which negotiates a common representation from heterogeneous agents' local features via a negotiator, with sender-receiver pairs and multi-dimensional alignment losses, for immutable heterogeneous collaborative perception.
Problem
Collaborative perception among agents with different fixed perception models creates domain gaps in shared intermediate features, degrading fusion performance; prior common-representation methods designate one agent's representation as the shared space, making alignment hard for dissimilar agents.
Approach
Instead of designating one agent's representation as the common space, a training-time negotiator fuses local representations from all modalities (FPN with per-level importance estimators) into a common representation; senders map local features into it, receivers map back, supervised by cyclic distribution loss and multi-dimensional alignment (distribution, structural keypoint-similarity, and occ
Key results
- On OPV2V-H, NegoCollab reaches AP@0.5 of 0.745 across all agent types versus 0.676 for NegoCollab-P and 0.512 for MPDA, with AP@0.7 of 0.555.
- On V2V4Real and DAIR-V2X, NegoCollab achieves AP@0.5 of 0.605 and 0.397, outperforming the best common-representation baselines.
- Compared with directly designating the protocol agent's representation as common space, KL-measured domain gap is reduced by about 93.5% on average.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use negotiated common representations instead of one-to-one adaptation | NegoCollab · OPV2V-H, All agent types | AP@0.5: 0.512 -> 0.745 | Helped |
| use negotiated common representations instead of one-to-one adaptation | NegoCollab · OPV2V-H, All agent types | AP@0.7: 0.435 -> 0.555 | Helped |
| use negotiated common representations instead of one-to-one adaptation | NegoCollab · OPV2V-H, m1m2 | AP@0.5: 0.865 -> 0.872 | Helped |
| use negotiated common representations instead of one-to-one adaptation | NegoCollab · V2V4Real, m1m3 | AP@0.5: 0.613 -> 0.605 | Hurt |
| use negotiated common representations instead of one-to-one adaptation | NegoCollab · DAIR-V2X, m1m2 | AP@0.5: 0.443 -> 0.397 | Hurt |
| add a negotiator to derive the common representation (ablation) | NegoCollab training · OPV2V-H, m1m2, before task adaptation | AP@0.5: 0.617 -> 0.609 | Helped |
| add structural alignment loss (ablation) | NegoCollab training with negotiator · OPV2V-H, m1m2, before task adaptation | AP@0.5: 0.609 -> 0.655 | Helped |
| add pragmatic alignment loss (ablation) | NegoCollab training with negotiator · OPV2V-H, m1m2, before task adaptation | AP@0.5: 0.609 -> 0.671 | Helped |
| add both structural and pragmatic alignment losses (ablation) | NegoCollab training with negotiator · OPV2V-H, m1m2, before task adaptation | AP@0.5: 0.609 -> 0.711 | Helped |
| replace Convext sender recombiner with ResMlp (ablation) | NegoCollab · OPV2V-H, m1m2, before task adaptation | AP@0.5: 0.711 -> 0.633 | Hurt |
| replace Fused Axial Attention aligner with ResMlp (ablation) | NegoCollab · OPV2V-H, m1m2, before task adaptation | AP@0.5: 0.711 -> 0.697 | Hurt |
| replace FPN negotiator with ResMlp (ablation) | NegoCollab · OPV2V-H, m1m2, before task adaptation | AP@0.5: 0.711 -> 0.705 | Hurt |
Surprising findings
- Under localization noise, NegoCollab outperforms late fusion by large margins at higher noise, while late fusion wins without noise.
- Common representations negotiated from agents with better perception encoders generalize better to new agents.
Limitations
- Once the common representation is negotiated it is fixed, so new agents aligned to it incur information loss.
- Collaboration with new agents m3 and m4 is slightly below one-to-one adaptation methods in some scenarios.
Experimental setup
Datasets: OPV2V-H V2V4Real DAIR-V2X
Benchmarks: OPV2V-H V2V4Real DAIR-V2X
Metrics: AP@0.5 AP@0.7 KL divergence
Baselines: MPDA PnPDA STAMP MPDA-P PnPDA-P NegoCollab-P Late Fusion
Builds on: MPDA PnPDA STAMP ConvNeXt focal loss
Closest papers by meaning
- An Extensible Framework for Open Heterogeneous Collaborative Perception (ICLR 2024 · similarity 0.91)
- Pragmatic Heterogeneous Collaborative Perception via Generative Communication Mechanism (NeurIPS 2025 · similarity 0.90)
- CoPEFT: Fast Adaptation Framework for Multi-Agent Collaborative Perception with Parameter-Efficient Fine-Tuning (AAAI 2025 · similarity 0.89)
- What Makes Good Collaborative Views? Contrastive Mutual Information Maximization for Multi-Agent Perception (AAAI 2024 · similarity 0.88)
- One is Plenty: A Polymorphic Feature Interpreter for Immutable Heterogeneous Collaborative Perception (CVPR 2025 · similarity 0.88)
- Communication-Efficient Collaborative Perception via Information Filling with Codebook (CVPR 2024 · similarity 0.88)
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.