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

MAPLE: Latent Multi-Agent Play for End-to-End Autonomous Driving

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.

Proposes MAPLE, a simulator-free framework that performs reactive multi-agent autoregressive rollouts in a VLA model's latent space for closed-loop SFT and GRPO-based RL with global, agent-specific, and diversity rewards, reaching state-of-the-art Bench2Drive driving score.

Problem

End-to-end VLA driving planners are trained open-loop on logged data, suffering covariate shift and ignoring reactive multi-agent interactions, while existing closed-loop RL needs symbolic or pixel-level simulators that are costly or low-fidelity.

Approach

Roll out ego and reactive agent states autoregressively in the VLA's latent token space (no simulator or image generation), supervise with SFT on those rollouts, then apply GRPO with global, per-agent, and behavior-diversity rewards over multiple planner heads.

Key results

• MAPLE reaches Driving Score 85.2 and SR 67.1 on Bench2Drive with NC conditioning; the PID-controller variant reaches DS 88.3 and SR 70.5, versus ORION at DS 77.7 and GeRo at 81.9.
• Full rollout across pretraining, SFT, and RL improves DS by 12.3 points over the baseline (68.1 to 80.4 for ego-only RL).
• Multi-agent play with diversity regularization reaches DS 85.2 and SR 67.5 in the ablation, above the single-planner self-play variant (DS 84.1).

Surprising findings

• Reactive-agent rollout with diversity regularization gives gains over ego-only rollout, but the closed-loop gains are modest (about 2.6 DS from 0 to 8 reactive agents).

Limitations

• Multi-agent interactions are modeled in latent space rather than physical simulation, which may limit fidelity for high-precision dynamics.
• Experiments are limited to Bench2Drive (CARLA) with no real-world data evaluation.

Experimental setup

Datasets: Bench2Drive

Benchmarks: Bench2Drive

Metrics: Driving Score Success Rate Efficiency Comfort Multi-Ability Avg L2

Baselines: SimLingo HiP-AD ORION ReCogDrive DiffRefiner GeRo DriveTransformer-L VAD UniAD-Base TCP-traj ThinkTwice DriveAdapter

Builds on: GRPO VAD ORION Bench2Drive

Closest papers by meaning

• Generative Scenario Rollouts for End-to-End Autonomous Driving (NeurIPS 2026 · similarity 0.89)
• DriveAgent-R1: Advancing VLM-based Autonomous Driving with Active Perception and Hybrid Thinking (ICLR 2026 · similarity 0.88)
• LaST-VLA: Thinking in Latent Spatio-Temporal Space for Vision-Language-Action in Autonomous Driving (NeurIPS 2026 · similarity 0.87)
• AutoVLA: A Vision-Language-Action Model for End-to-End Autonomous Driving with Adaptive Reasoning and Reinforcement Fine-Tuning (NeurIPS 2025 · similarity 0.87)
• Model-Based Policy Adaptation for Closed-Loop End-to-end Autonomous Driving (NeurIPS 2025 · similarity 0.87)
• Prioritizing Perception-Guided Self-Supervision: A New Paradigm for Causal Modeling in End-to-End Autonomous Driving (NeurIPS 2025 · similarity 0.87)