Research map / Multi-agent systems and games / Human-AI and goal-conditioned
LASIL: Learner-Aware Supervised Imitation Learning For Long-term Microscopic Traffic Simulation
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 LASIL, a supervised imitation learning method that uses a context-conditioned VAE to augment expert states with learner-state awareness, enabling stable long-term microscopic traffic simulation on pNEUMA.
Problem
Behavior-cloning-based and adversarial imitation learning traffic simulators suffer from covariate shift and fail to produce stable long-term microscopic and macroscopic simulations of human driving.
Approach
Augment expert states with a context-conditioned VAE trained jointly on expert and learner state distributions so augmented states are learner-aware, then train an EGAT policy via supervised NLL on real future trajectories, with LQR smoothing and on-road projection at rollout.
Key results
- On 20-second microscopic evaluation, LASIL achieves position RMSE 19.21 m and velocity RMSE 3.02 m/s, versus 29.78 m for the best baseline (MARL+BC).
- On 800-second macroscopic evaluation, LASIL reaches road speed RMSE 3.17 m/s and off-road rate 0.34%, versus 3.89 m/s and 2.92% for RAIL.
- Claims the first IL-based traffic simulator reproducing stable microscopic simulation for over 10 minutes, a 40x length improvement over prior state of the art.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add VAE augmentation (ablation) | LASIL · pNEUMA, 20 seconds | Position RMSE (m): 21.62±0.54 -> 19.21±0.44 | Helped |
| Add VAE augmentation (ablation) | LASIL · pNEUMA, 800 seconds | Road Density RMSE (veh/km): 55.65±0.44 -> 45.13±0.25 | Helped |
| Use a context-conditioned VAE instead of a naive VAE (ablation) | LASIL · pNEUMA, 20 seconds | Position RMSE (m): 23.03±0.87 -> 19.21±0.44 | Helped |
| Use a context-conditioned VAE instead of a naive VAE (ablation) | LASIL · pNEUMA, 800 seconds | Road Speed RMSE (m/s): 4.48±0.19 -> 3.17±0.14 | Helped |
| Add on-road projection (ablation) | LASIL · pNEUMA, 20 seconds | Off-road (%): 1.74±0.60 -> 0.28±0.01 | Helped |
| Add on-road projection (ablation) | LASIL · pNEUMA, 800 seconds | Off-road (%): 11.33±2.37 -> 0.34±0.03 | Helped |
| Add LQR smoothing (ablation) | LASIL · pNEUMA, 20 seconds | Velocity RMSE (m/s): 4.05±0.15 -> 3.02±0.07 | Helped |
| Add LQR smoothing (ablation) | LASIL · pNEUMA, 800 seconds | Road Density RMSE (veh/km): 61.68±0.96 -> 45.13±0.25 | Helped |
Surprising findings
- Removing LQR smoothing increases short-term off-road rate (0.22% vs 0.28% for full model is lower, but the authors note LQR raises short-term off-road rate relative to ablation), and removing it degrades other metrics.
- Plain behavior cloning yields a 33.43% off-road rate in short-term simulation, far higher than the adversarial and RL-based baselines.
Limitations
- Skips collision removal and does not report collision rate because vehicle size and heading information are unavailable in the dataset.
- Short-term off-road rate is higher with LQR smoothing than without, a trade-off noted by the authors.
- Evaluated only on a single urban dataset (pNEUMA, Athens).
Experimental setup
Datasets: pNEUMA
Benchmarks: pNEUMA
Metrics: position RMSE velocity RMSE minADE20 off-road rate road density RMSE road speed RMSE
Baselines: SUMO BC MARL MARL+BC PS-GAIL RAIL
Builds on: DAgger DART GAIL PS-GAIL edge-enhanced graph attention network
Closest papers by meaning
- Flow Matching-Based Autonomous Driving Planning with Advanced Interactive Behavior Modeling (NeurIPS 2025 · similarity 0.87)
- BehaviorGPT: Smart Agent Simulation for Autonomous Driving with Next-Patch Prediction (NeurIPS 2024 · similarity 0.87)
- Advancing Multi-agent Traffic Simulation via R1-Style Reinforcement Fine-Tuning (ICLR 2026 · similarity 0.86)
- Closed-Loop Supervised Fine-Tuning of Tokenized Traffic Models (CVPR 2025 · similarity 0.86)
- Trajeglish: Traffic Modeling as Next-Token Prediction (ICLR 2024 · similarity 0.86)
- Scenario Dreamer: Vectorized Latent Diffusion for Generating Driving Simulation Environments (CVPR 2025 · 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.