DynVLA: Learning World Dynamics for Action Reasoning in Autonomous Driving

We propose DynVLA, a driving VLA model that introduces a new CoT paradigm termed Dynamics CoT. DynVLA forecasts compact world dynamics before action generation, enabling more informed and physically grounded decision-making. To obtain compact dynamics representations, DynVLA introduces a Dynamics Tokenizer that compresses future evolution into a small set of dynamics tokens. Considering the rich environment dynamics in interaction-intensive driving scenarios, DynVLA decouples ego-centric and environment-centric dynamics, yielding more accurate world dynamics modeling. We then train DynVLA to generate dynamics tokens before actions through SFT and RFT, improving decision quality while maintaining latency-efficient inference. Compared to Textual CoT, which lacks fine-grained spatiotemporal understanding, and Visual CoT, which introduces substantial redundancy due to dense image prediction, Dynamics CoT captures the evolution of the world in a compact, interpretable, and efficient form. Extensive experiments on NAVSIM, Bench2Drive, and a large-scale in-house dataset demonstrate that DynVLA consistently outperforms Textual CoT and Visual CoT methods, validating the effectiveness and practical value of Dynamics CoT.

Safe autonomous driving requires more than recognizing what is in a camera image: the car also needs to anticipate what may happen next. For example, it should understand whether a vehicle ahead is likely to stop, whether another lane may open, or whether an upcoming curb limits where it can go. In this paper, we introduce DynVLA, a driving model that first predicts a compact description of how the surrounding scene is likely to change, and then uses this prediction to plan its next movement. Unlike methods that explain driving decisions only with text or by generating full future images, our approach focuses on the changes that matter for driving, making the reasoning faster and less cluttered. It also separates changes caused by the self-driving car from changes caused by other road users, making the prediction easier to use for planning. Tests on public driving benchmarks and a large internal dataset show that DynVLA makes better and safer driving decisions than previous reasoning-based methods, while keeping computation efficient.