Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Multi-agent Collaborative Perception via Motion-aware Robust Communication Network
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 MRCNet, a motion-aware two-stage communication network for multi-agent LiDAR collaborative 3D detection that handles pose noise via multi-scale robust fusion and motion blur via a GRU-style gated motion unit.
Problem
Collaborative perception systems assume ideal communication and perception, ignoring real-world noise such as pose noise, perception noise, and motion blur from moving objects, which degrades fusion and detection.
Approach
Combines a semantic-confidence-based feature selection with two-stage communication (multi-scale enhanced aggregation using pyramid sampling and multi-scale deformable self-attention, then collaborative feature aggregation) for pose-noise robustness, plus a gated motion unit that uses BEV feature differences over a memory bank of historical frames to refine the ego feature.
Key results
- On V2XSim under pose noise (0.2 m, 0.2 deg), MRCNet reaches AP0.7 69.82 vs. 66.24 for AttFuse and 65.06 for CoBEVT.
- On OPV2V, MRCNet achieves AP0.7 76.12 vs. 75.70 for V2X-ViT; on V2XSet, AP0.7 66.31 vs. 66.27 for AttFuse.
- Under increasing localization noise, baselines such as Where2comm and V2X-ViT degrade rapidly while MRCNet maintains superior AP0.7 (per Figure 4).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MRCNet instead of the strongest baseline | Collaborative 3D detection · V2XSim | AP0.5: 84.94 -> 85.33 | Helped |
| Use MRCNet instead of the strongest baseline | Collaborative 3D detection · V2XSim | AP0.7: 66.24 -> 69.82 | Helped |
| Use MRCNet instead of the strongest baseline | Collaborative 3D detection · OPV2V | AP0.5: 89.27 -> 89.77 | Helped |
| Use MRCNet instead of the strongest baseline | Collaborative 3D detection · OPV2V | AP0.7: 75.70 -> 76.12 | Helped |
| Use MRCNet instead of the strongest baseline | Collaborative 3D detection · V2XSet | AP0.5: 84.84 -> 85.00 | Helped |
| Use MRCNet instead of the strongest baseline | Collaborative 3D detection · V2XSet | AP0.7: 66.27 -> 66.31 | Helped |
| Add MEM (ablation) | MRCNet · V2XSim | AP@0.7: 64.28 -> 67.64 | Helped |
| Add MEM (ablation) | MRCNet · OPV2V | AP@0.7: 74.01 -> 74.63 | Helped |
| Add MEM (ablation) | MRCNet · V2XSet | AP@0.7: 63.88 -> 64.71 | Helped |
| Add feature selection (ablation) | MRCNet with MEM · V2XSim | AP@0.7: 67.64 -> 69.15 | Helped |
| Add feature selection (ablation) | MRCNet with MEM · OPV2V | AP@0.7: 74.63 -> 75.31 | Helped |
| Add feature selection (ablation) | MRCNet with MEM · V2XSet | AP@0.7: 64.71 -> 65.31 | Helped |
Surprising findings
- Early fusion outperforms several intermediate fusion methods on V2XSim AP0.5 (84.94 vs. 81-83 for many intermediate baselines).
- Where2comm's performance degrades as transmission volume increases because it struggles to filter noisy data.
Limitations
- Evaluated only on simulated datasets (V2XSim, OPV2V, V2XSet) rather than real-world data.
- Performance saturates or degrades beyond three history frames and excessive misaligned feature transmission degrades results.
- Communication issues and model heterogeneity are left as future work.
Experimental setup
Datasets: V2XSim OPV2V V2XSet
Benchmarks: V2XSim OPV2V V2XSet
Metrics: AP@0.5 AP@0.7
Baselines: No Fusion Late Fusion Early Fusion When2com V2VNet AttFuse V2X-ViT DiscoNet CoBEVT Where2comm
Builds on: PointPillars Where2comm V2X-ViT deformable attention GRU
Closest papers by meaning
- CoSDH: Communication-Efficient Collaborative Perception via Supply-Demand Awareness and Intermediate-Late Hybridization (CVPR 2025 · similarity 0.88)
- DSRC: Learning Density-Insensitive and Semantic-Aware Collaborative Representation Against Corruptions (AAAI 2025 · similarity 0.86)
- What Makes Good Collaborative Views? Contrastive Mutual Information Maximization for Multi-Agent Perception (AAAI 2024 · similarity 0.86)
- An Extensible Framework for Open Heterogeneous Collaborative Perception (ICLR 2024 · similarity 0.86)
- RCP-Bench: Benchmarking Robustness for Collaborative Perception Under Diverse Corruptions (CVPR 2025 · similarity 0.85)
- RCDN: Towards Robust Camera-Insensitivity Collaborative Perception via Dynamic Feature-based 3D Neural Modeling (NeurIPS 2024 · similarity 0.85)
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.