On-Policy Distillation of Language Models for Autonomous Vehicle Motion Planning
A training method shrinks LLM-based autonomous vehicle planners 5x while nearly matching teacher performance — but only on a single benchmark.
The Thesis
Language models can plan driving routes by converting a vehicle's surroundings into text and generating a sequence of waypoints — effectively treating motion planning as a writing problem. The catch is that capable models are too large to run inside a car's onboard computer. This paper tests whether a smaller 'student' model can learn to plan nearly as well as a larger 'teacher' model through a technique called on-policy generalized knowledge distillation (GKD), where the student trains on its own generated outputs while receiving detailed token-by-token feedback from the teacher. On the nuScenes benchmark — a standard dataset of real urban driving scenes — the 5x-smaller student nearly matched teacher performance and significantly outperformed a reinforcement-learning baseline trained with similar signals. If these results generalize beyond the benchmark, GKD offers a practical path to deploying LLM-based planners without requiring expensive onboard hardware.
Catalyst
LLM-based planning frameworks like GPT-Driver only recently demonstrated that driving scenes could be represented as language prompts at useful fidelity. Distillation techniques like GKD have matured in parallel with the LLM scaling wave, making it timely to combine them. Meanwhile, automotive OEMs are under real cost pressure to shrink compute hardware in vehicles, creating immediate practical demand for exactly this kind of model compression.
What's New
The GPT-Driver framework, on which this paper builds, introduced the idea of encoding driving scenes as text and generating trajectories with chain-of-thought reasoning (step-by-step explanations produced before the final answer). Prior distillation methods typically trained a student on pre-collected teacher outputs — called offline distillation — which means the student never sees its own mistakes. This paper applies on-policy distillation, where the student generates its own trajectories during training and receives dense, per-token probability feedback from the teacher, allowing it to learn from its own errors rather than just mimicking a fixed dataset. The authors claim this on-policy feedback loop is the key reason GKD outperforms a reinforcement-learning approach that uses teacher probabilities as reward signals.
The Counter
This paper evaluates on a single benchmark — nuScenes — which consists of pre-recorded urban driving clips, not closed-loop real-world driving. Performing well on trajectory prediction in nuScenes does not prove the system is safe or effective when the car's own actions change what happens next. The 5x size reduction sounds impressive, but the paper does not specify what absolute model sizes are involved; a smaller-but-still-large model may not fit in a real vehicle's compute envelope. Chain-of-thought reasoning also adds latency at inference time, which matters when a car needs to react in milliseconds — and this is not addressed. Finally, the comparison baseline is an RL approach using the same teacher signals, not the strongest existing compact AV planners, so it is unclear how GKD stacks up against purpose-built small models trained on classical supervision. The result is a useful research finding, but the gap between a nuScenes score and a deployable product remains very wide.
Longs
- MBLY (Mobileye) — direct play on onboard AV compute and planning software
- APTV (Aptiv) — AV software stack supplier to major OEMs
- INTC — Intel's AV division and edge inference silicon
- QUIK (QuickLogic) — low-power edge AI inference chips relevant to constrained onboard systems
- KTOS (Kratos Defense) — autonomous systems platforms where compact AI planning matters
Shorts
- Suppliers of large onboard compute modules (e.g., NVIDIA DRIVE AGX at full spec) — if smaller models suffice, OEMs may buy cheaper hardware tiers
- Companies selling rule-based or classical motion planning middleware — LLM planners with strong small-model performance erode the 'LLMs are too big for cars' objection that protects incumbents
Enablers (Picks & Shovels)
- nuScenes dataset (open benchmark from Motional used in this paper's experiments)
- GPT-Driver framework (open research framework this work builds directly upon)
- Hugging Face Transformers (open-source library for fine-tuning and distilling LLMs)
- NVIDIA DRIVE platform (onboard inference hardware where compressed planners would ultimately run)
- vLLM / TensorRT-LLM (inference optimization tools that complement model size reduction)
Private Watchlist
- Wayve — LLM-integrated end-to-end AV planning
- Waabi — generative AI for autonomous trucking
- Ghost Autonomy — LLM-based highway driving stack
- Comma.ai — open-source onboard AV software with constrained hardware focus
Resources
The Paper
Large language models (LLMs) have recently demonstrated strong potential for autonomous vehicle motion planning by reformulating trajectory prediction as a language generation problem. However, deploying capable LLMs in resource-constrained onboard systems remains a fundamental challenge. In this paper, we study how to effectively transfer motion planning knowledge from a large teacher LLM to a smaller, more deployable student model. We build on the GPT-Driver framework, which represents driving scenes as language prompts and generates waypoint trajectories with chain-of-thought reasoning, and investigate two student training paradigms: (i) on-policy generalized knowledge distillation (GKD), which trains the student on its own self-generated outputs using dense token-level feedback from the teacher, and (ii) a dense-feedback reinforcement learning (RL) baseline that uses the teacher's log-probabilities as per-token reward signals in a policy gradient framework. Experiments on the nuScenes benchmark show that GKD substantially outperforms the RL baseline and closely approaches teacher-level performance despite a 5$\times$ reduction in model size. These results highlight the practical value of on-policy distillation as a principled and effective approach to deploying LLM-based planners in autonomous driving systems.