Research map / Large language models / Agents, code and math
Language-Driven Interactive Traffic Trajectory Generation
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 InteractTraj, a language-driven traffic trajectory generator that uses an LLM (GPT-4) to convert descriptions into interaction-aware numerical codes and a code-to-trajectory decoder with two-step map-vehicle-interaction attention to generate interactive multi-vehicle trajectories.
Problem
Existing language-driven traffic trajectory generators (e.g., LCTGen) generate each traffic participant independently, failing to model interactions between vehicles such as overtaking, merging, yielding, and following.
Approach
Bridge abstract language and trajectories with discrete interaction-aware numerical codes (relative direction/distance over time, vehicle trend/action codes, map codes), and fuse map, vehicle, and interaction features via two-step cross-attention in the decoder.
Key results
- On WOMD, InteractTraj reaches mADE 1.067 and mFDE 2.190 vs LCTGen 1.262 and 2.696, a 15.4%/18.7% average improvement claimed.
- On nuPlan, InteractTraj achieves mADE 0.962 and mFDE 1.987 vs LCTGen 1.161 and 2.497.
- User study: 73.7% average preference for InteractTraj over LCTGen; 72.4% positive interaction-fulfillment vs 31.5% for LCTGen.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use InteractTraj instead of LCTGen | language-conditioned trajectory reconstruction · WOMD | mADE: 1.262 -> 1.067 | Helped |
| Use InteractTraj instead of LCTGen | language-conditioned trajectory reconstruction · nuPlan | mADE: 1.161 -> 0.962 | Helped |
| Add interaction codes (IC) (ablation) | InteractTraj · WOMD | mADE: 1.205 -> 1.067 | Helped |
| Add relative distance (RD) to interaction codes (ablation) | InteractTraj · WOMD | mADE: 1.167 -> 1.067 | Helped |
| Add relative position (RP) to interaction codes (ablation) | InteractTraj · WOMD | mADE: 1.194 -> 1.067 | Helped |
| Add relative distance loss (ablation) | InteractTraj · WOMD | mADE: 1.165 -> 1.067 | Helped |
| Set interaction-code distance gap to 10 instead of 15 (ablation) | InteractTraj · WOMD | mADE: 1.067 -> 1.087 | Hurt |
| Set interaction-code distance gap to 20 instead of 15 (ablation) | InteractTraj · WOMD | mADE: 1.067 -> 1.117 | Hurt |
| Use 4 interaction-code areas instead of 6 (ablation) | InteractTraj · WOMD | mADE: 1.067 -> 1.237 | Hurt |
| Use 8 interaction-code areas instead of 6 (ablation) | InteractTraj · WOMD | mADE: 1.067 -> 1.069 | Hurt |
Surprising findings
- The interaction-code ablation without relative-position codes still improves over LCTGen, suggesting vehicle-code design alone contributes a large share of gains.
Limitations
- Only generates vehicle trajectories, not other traffic participants.
- Map generation is limited by the pre-defined map library.
- Evaluation uses reconstruction from ground-truth-derived codes rather than true language-trajectory pairs.
Experimental setup
Datasets: Waymo Open Motion Dataset nuPlan
Benchmarks: Waymo Open Motion Dataset nuPlan
Metrics: mADE minADE mFDE minFDE SCR Hausdorff distance user preference
Baselines: TrafficGen LCTGen
Builds on: LCTGen multi-context gating TrafficGen
Closest papers by meaning
- Trajeglish: Traffic Modeling as Next-Token Prediction (ICLR 2024 · similarity 0.89)
- Flow Matching-Based Autonomous Driving Planning with Advanced Interactive Behavior Modeling (NeurIPS 2025 · similarity 0.86)
- TrajMamba: An Efficient and Semantic-rich Vehicle Trajectory Pre-training Model (NeurIPS 2025 · similarity 0.85)
- SOLVE: Synergy of Language-Vision and End-to-End Networks for Autonomous Driving (CVPR 2025 · similarity 0.85)
- WOMD-Reasoning: A Large-Scale Dataset for Interaction Reasoning in Driving (ICML 2025 · similarity 0.85)
- Less Language, More Latents: Annotation-Efficient VLAs for Driving (NeurIPS 2026 · 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.