Research map / Probabilistic and scientific modeling / Molecules and proteins
NEST: A Neuromodulated Small-world Hypergraph Trajectory Prediction Model for Autonomous Driving
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 NEST, a trajectory predictor that builds interaction hypergraphs with a Newman-Watts small-world hyperedge generator whose thresholds are set by a neuromodulator, and reports SOTA on nuScenes, MoCAD and HighD.
Problem
Trajectory prediction for autonomous driving struggles to model dynamic, group-wise and long-range agent interactions efficiently in dense heterogeneous traffic.
Approach
Combines a Newman-Watts small-world network used to generate hyperedges with a neuromodulator that adaptively sets the clustering threshold and connection probability from agent features, followed by vertex-hyperedge message passing and lane-context attention fusion with a multimodal Laplace-distribution predictor.
Key results
- On nuScenes, NEST reaches minADE_5 of 1.18 versus 1.38 for the best listed prior method (SeFlow), a 14.5% improvement.
- On HighD, NEST achieves RMSE 0.48 at 5 s versus 0.62 for BAT, and reports 27.3% and 22.6% gains over BAT at 4 s and 5 s.
- NEST runs in 11.6 ms per 12 agents on an RTX 3090, faster than all listed baselines (e.g., 38 ms for Trajectron++, 215 ms for PGP).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the Neuromodulator (ablation) | NEST · nuScenes | mADE_5: 1.21 -> 1.18 | Helped |
| add the Neuromodulated Small-world Network (ablation) | NEST · nuScenes | mADE_5: 1.25 -> 1.18 | Helped |
| use hypergraph learning instead of a conventional graph neural network (ablation) | NEST · nuScenes | mADE_5: 1.48 -> 1.18 | Helped |
| add Context Fusion (ablation) | NEST · nuScenes | mADE_5: 1.37 -> 1.18 | Helped |
| add the Multi-modal Decoder (ablation) | NEST · nuScenes | mADE_5: 1.22 -> 1.18 | Helped |
| use NEST instead of BAT | trajectory prediction · MoCAD, 5 s | RMSE: 2.88 -> 2.42 | Helped |
| use NEST instead of BAT | trajectory prediction · HighD, 4 s | RMSE: 0.44 -> 0.32 | Helped |
| use NEST instead of BAT | trajectory prediction · HighD, 5 s | RMSE: 0.62 -> 0.48 | Helped |
| use NEST instead of BAT | trajectory prediction · MoCAD, 2 s | RMSE: 0.74 -> 0.75 | Hurt |
| use NEST instead of Trajectron++ | trajectory prediction · nuScenes | inference time (ms): 38 -> 11.6 | Helped |
Surprising findings
- NEST is slightly worse than BAT at 2 s on MoCAD (0.75 vs 0.74 RMSE), which the authors acknowledge.
Limitations
- Inference time comparison uses baseline timings taken from VisionTrap on a different GPU (RTX 3090 Ti), so the comparison is not hardware-matched.
- Does not beat BAT at the 2 s horizon on MoCAD.
- Hyperparameters and training details are not described in the provided text.
Experimental setup
Datasets: nuScenes MoCAD HighD
Benchmarks: nuScenes MoCAD HighD
Metrics: minADE_5 minADE_1 minFDE_1 RMSE inference time (ms)
Baselines: DLow-AF Trajectron++ MultiPath LDS-AF AgentFormer LaPred STGM GoHome ContextVAE EMSIN SeFlow CS-LSTM
Builds on: Newman-Watts small-world model hypergraph neural networks AgentFormer Gumbel-softmax
Closest papers by meaning
- LaKD: Length-agnostic Knowledge Distillation for Trajectory Prediction with Any Length Observations (NeurIPS 2024 · similarity 0.88)
- BAT: Behavior-Aware Human-Like Trajectory Prediction for Autonomous Driving (AAAI 2024 · similarity 0.88)
- Trajeglish: Traffic Modeling as Next-Token Prediction (ICLR 2024 · similarity 0.87)
- DriveGPT: Scaling Autoregressive Behavior Models for Driving (ICML 2025 · similarity 0.86)
- Hybrid-Driving: An Autonomous Driving Decision Framework Integrating Large Language Models, Knowledge Graphs and Driving Rules (AAAI 2025 · similarity 0.86)
- DeMo: Decoupling Motion Forecasting into Directional Intentions and Dynamic States (NeurIPS 2024 · similarity 0.86)
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.