Research map / Multi-agent systems and games / Human-AI and goal-conditioned

Learning to Drive in New Cities Without Human Demonstrations

NeurIPS 2026 · Multi-agent systems and games › Human-AI and goal-conditioned · approach: Data-driven traffic simulation · 0 citations (OpenAlex)

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.

Adapts a Boston-trained driving policy to new cities (e.g., Singapore) via KL-regularized self-play MARL in a map-based GPUDrive simulator built from target-city maps and meta-information, with no target-city human demonstrations.

Problem

Adapting autonomous driving policies to new cities normally requires collecting costly human demonstrations in each target city, because road geometry, traffic rules, and interaction patterns shift across cities.

Approach

Construct a target-city simulator from the lane-level map and coarse meta-information (speed limits, traffic density), then adapt an imitation-learned source policy via self-play multi-agent PPO with a minimal success/collision/offroad reward, regularized by a reverse KL to the source prior.

Key results

• Boston-to-Singapore transfer: NOMAD raises success rate from 42.25% (zero-shot BC) to a frontier of 87.28% to 94.27%, with realism 0.7570 to 0.7697 versus 0.6795 for zero-shot.
• NOMAD Pareto frontier dominates BC trained on 3,200 Singapore scenarios (55.49% success, 0.7011 realism) and matches much of the gain of target-demonstration training.
• Removing the imitation prior (RL from scratch in target city) yields high success but lower realism, with kinematic realism failing to improve.

Surprising findings

• The success-realism trade-off is mild: a 7-point success gain costs only about 0.012 realism, so selecting the highest-success checkpoint is practical without target trajectories.
• Self-play in the source city (Boston) yields policies that transfer poorly to Singapore, showing target-city map grounding is necessary.

Limitations

• Realism is measured with WOSAC on nuPlan logged trajectories, which are used only for evaluation; the realism metric is city-dependent and not comparable across cities.
• Differences in driving culture and social conventions are not captured by map geometry alone.
• Scenario generation relies on heuristic placement and city-specific density fitted from logged nuPlan data.

Experimental setup

Datasets: nuPlan

Benchmarks: nuPlan (Boston to Singapore, Singapore to Boston, Singapore to Pittsburgh)

Metrics: success rate WOSAC realism meta score collision rate offroad rate ADE

Baselines: Behavior cloning (source) Behavior cloning (target) RL from scratch (target) BC + RL (source) BC + RL (target) Random Constant velocity Expert oracle

Builds on: PPO IPPO GPUDrive WOSAC nuPlan

Closest papers by meaning

• IntersectionZoo: Eco-driving for Benchmarking Multi-Agent Contextual Reinforcement Learning (ICLR 2025 · similarity 0.86)
• Beyond Self-Play and Scale: A Behavior Benchmark for Generalization in Autonomous Driving (NeurIPS 2026 · similarity 0.86)
• Robust Autonomy Emerges from Self-Play (ICML 2025 · similarity 0.86)
• MAPLE: Latent Multi-Agent Play for End-to-End Autonomous Driving (NeurIPS 2026 · similarity 0.86)
• Diffusion-Based Planning for Autonomous Driving with Flexible Guidance (ICLR 2025 · similarity 0.85)
• World Models as Adversaries: Multi-Agent Self-Play Fine-Tuning for Robust Motion Planning (NeurIPS 2026 · similarity 0.85)