Research map / Multi-agent systems and games / Cooperative multi-agent RL
TurboTrain: Towards Efficient and Balanced Multi-Task Learning for Multi-Agent Perception and Prediction
Zewei Zhou, Seth Z. Zhao, Tianhui Cai, Zhiyu Huang, Bolei Zhou, Jiaqi Ma
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add multi-agent spatiotemporal pretraining (ablation) | V2XPnP · V2XPnP-Seq-VC | EPA: 15.0 -> 43.4 | Improved |
| include multi-agent fusion module in pretraining (ablation) | TurboTrain pretraining · V2XPnP-Seq-VC | EPA: 40.8 -> 43.4 | Improved |
| use both point and occupancy reconstruction losses (ablation) | TurboTrain pretraining · V2XPnP-Seq-VC | EPA: 40.2/40.8 -> 43.4 | Improved |
| add conflict-suppressing balance stage (ablation) | V2XPnP pretrained · V2XPnP-Seq-VC | EPA: 43.4 -> 45.5 | Improved |
| replace conflict-suppression with GradNorm (ablation) | TurboTrain Balance · V2XPnP-Seq-VC | EPA: 43.4 -> 34.2 | Hurt |
| balance at every gradient step instead of hybrid free/balanced (ablation) | TurboTrain Balance · V2XPnP-Seq-VC | AP@0.5: 72.2 -> 0.0 | Hurt |
| shorten free/balance interval to (1000,500) (ablation) | TurboTrain Balance · V2XPnP-Seq-VC | EPA: 45.5 -> 38.1 | Hurt |
| lengthen free/balance interval to (3000,1500) (ablation) | TurboTrain Balance · V2XPnP-Seq-VC | EPA: 45.5 -> 42.8 | Hurt |
| use all historical frames in reconstruction instead of half (ablation) | TurboTrain pretraining · V2XPnP-Seq-VC | EPA: 38.8 -> 43.4 | Improved |
| raise mask ratio from 0.7 to 0.9 (ablation) | TurboTrain pretraining · V2XPnP-Seq-VC | EPA: 43.4 -> 42.1 | Hurt |
| replace manual multi-stage training with TurboTrain | V2XPnP · V2XPnP-Seq-VC | EPA: 41.8 -> 45.5 | Improved |
| replace manual training with TurboTrain on F-Cooper | F-Cooper · V2XPnP-Seq-VC | AP@0.5: 66.5 -> 66.3 | No clear effect |
Closest papers by meaning
- CoPEFT: Fast Adaptation Framework for Multi-Agent Collaborative Perception with Parameter-Efficient Fine-Tuning (AAAI 2025 · similarity 0.85)
- MAPLE: Latent Multi-Agent Play for End-to-End Autonomous Driving (NeurIPS 2026 · similarity 0.84)
- An Extensible Framework for Open Heterogeneous Collaborative Perception (ICLR 2024 · similarity 0.84)
- Learning Generalizable Skills from Offline Multi-Task Data for Multi-Agent Cooperation (ICLR 2025 · similarity 0.84)
- MAT-Agent: Adaptive Multi-Agent Training Optimization (NeurIPS 2025 · similarity 0.83)
- One is Plenty: A Polymorphic Feature Interpreter for Immutable Heterogeneous Collaborative Perception (CVPR 2025 · similarity 0.83)
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.
