# Distill to Think, Foresee to Act: Cognitive-Physical Reinforcement Learning for Autonomous Driving

DevFeed: [Distill to Think, Foresee to Act: Cognitive-Physical Reinforcement Learning for Autonomous Driving](<https://devfeed.tech/articles/distill-to-think-foresee-to-act-cognitive-physical-reinforcement-learning-for-autonomous-driving-116710.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2605.21139>)

Author: Yang Wu, Qiang Meng, Zhaojiang Liu, Youquan Liu, Jian Yang, Jin Xie

Published: 2026-11-04T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [Reinforcement learning](<https://devfeed.tech/topics/reinforcement-learning.md>), [autonomous vehicles](<https://devfeed.tech/topics/autonomous-vehicles.md>), [World models](<https://devfeed.tech/topics/world-models.md>), [AI agents in autonomous vehicles](<https://devfeed.tech/topics/ai-agents-in-autonomous-vehicles.md>)

Tags: [autonomous-driving](<https://devfeed.tech/tags/autonomous-driving.md>), [cs-cv](<https://devfeed.tech/tags/cs-cv.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [reinforcement-learning](<https://devfeed.tech/tags/reinforcement-learning.md>), [safety](<https://devfeed.tech/tags/safety.md>), [world-model](<https://devfeed.tech/tags/world-model.md>)

## AI overview

CoPhy combines a distilled cognitive model with an action-conditioned physical world model to train autonomous driving policies using reinforcement learning. It distills vision-language model knowledge into a bird's-eye-view (BEV) encoder, then removes the VLM to retain the learned capability without inference cost; an optional language interface accepts human commands. An autoregressive BEV model predicts future semantic maps for candidate actions, providing rollouts from which safety metrics are derived. GRPO training uses a dual reward: physical rewards from BEV rollouts enforce safety constraints, while a language-aligned scorer rewards intent compliance. The abstract reports state-of-the-art results on NAVSIM v1 and v2, plus safer scene compliance and flexible control through user-defined language instructions; it gives no numerical results or implementation details.

## Source excerpt

arXiv:2605.21139v3 Announce Type: replace-cross Abstract: Current end-to-end autonomous driving models are fundamentally constrained by the behavioral cloning ceiling of imitation learning. While reinforcement learning offers a path to smarter autonomy, it demands two missing pieces of infrastructure: (1) a cognitive foundation that understands traffic semantics and driving intent, and (2) a foresighted physical environment that can anticipate the consequences of candidate actions. To this end, we propose CoPhy, a CognitivePhysical reinforcement learning framework for autonomous driving. To distill to think, we distill VLM knowledge into the BEV encoder and then discard the VLM entirely, retaining cognitive ability at zero inference cost while releasing the cognitive channel as a pluggable interface for optional human language commands. To foresee to act, we build an auto-regressive BEV world model that explicitly predicts future semantic maps conditioned on candidate actions, serving as an interpretable physical sandbox from which safety metrics are directly derived. Built upon this dual infrastructure, we optimize the driving policy via GRPO with a novel dual-reward mechanism: a physical reward derived from BEV rollouts enforces hard safety constraints, while a cognitive reward from a language-aligned scorer ensures intent compliance. Extensive experiments demonstrate that CoPhy not only achieves state-of-the-art results on NAVSIM v1 and v2 benchmarks, but also enables safer driving via cognitively informed scene compliance and flexible intent control through user-defined language instructions. Project page: https://wuyang98.github.io/CoPhy/.