.png){kind=icon}
.png){kind=icon}
Return
Multi-Agent Motion Forecasting via Mixed Supervision
.png){kind=icon}
PRE
AI
DOI：10.1109/LRA.2025.3625511.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Accurate motion forecasting is critical for safe decision-making in autonomous driving. Existing methods lack explicit supervision during the generation of multimodal candidate trajectories, which leads to inconsistent candidate quality and difficulty balancing trajectory diversity with scene adaptability. We propose Mix-Forecast, a motion forecasting framework based on mixed supervision that explicitly optimizes the trajectory decoding process by combining multi-hypothesis supervision with competitive optimal supervision. We design a mixed supervision decoding mechanism that fuses the layer-by-layer auxiliary guidance of multi-hypothesis supervision with competitive optimal supervision in a single decoder. This mechanism improves the diversity and accuracy of candidate trajectories without adding extra computational cost. We also construct the Mamba-Transformer hybrid architecture, which efficiently integrates historical motion features and spatiotemporal context through cross-layer memory transfer and selective spatiotemporal state space modeling. Furthermore, we represent high-dimensional trajectories in the reproducing kernel Hilbert space (RKHS) and design an exclusivity loss based on a regularized kernel Kullback-Leibler (KKL) divergence to enhance prediction diversity. Experiments on the Argoverse motion forecasting benchmarks show that Mix-Forecast performs exceptionally well, and ablation studies confirm the complementary contributions of mixed supervision, cross-layer memory, and KKL loss.
Keywords:
Mixed supervision
multi-agent motion forecasting
autonomous driving
Mamba hybrid architecture
Journal
I
IF:
5.3
Papers:
1.7K
Citations:
3.9W