NVIDIA squeezes a driving reasoning model onto onboard silicon: closed-loop RL lifts driving performance 48.9%, with 160–190 ms pixel-to-trajectory latency
Synopsis
NVIDIA released Alpamayo 2 Edge, an edge-oriented reasoning vision-language-action model that pairs a 2B-parameter Nemotron reasoning backbone with a diffusion action expert to produce Chain-of-Causation reasoning and future trajectories on DRIVE AGX Thor, generating reasoning only when needed. Closed-loop reinforcement learning post-training in AlpaGym improves closed-loop driving performance by 48.9% over the best-performing checkpoint without closed-loop RL, and the optimized model reaches 160–190 ms pixel-to-trajectory inference latency depending on the amount of reasoning.
Interpretation
An onboard driving model can now emit both causal reasoning and a driving trajectory inside the inference budget, instead of moving a large model into the vehicle unchanged. Larger models provide a foundation for these capabilities, but moving them into a vehicle requires reducing model size, preserving useful reasoning, and validating performance on the target hardware. After closed-loop reinforcement learning post-training in AlpaGym, closed-loop driving performance improves by 48.9% over the best-performing checkpoint without closed-loop RL.
The model generates reasoning only when needed and sometimes emits a special <NOMINAL> token instead of a reasoning trace during routine driving, conserving inference compute. This adaptive mechanism sits alongside offline distillation, multi-resolution surround-view inputs, DFlash2 speculative decoding, and quantization and TensorRT deployment recipes as the capacity-efficiency balance. On the reference configuration of DRIVE AGX Thor-X with DRIVE OS 7.2.7.0, CUDA 13.4, and TensorRT 11.3.1.9, the optimized model measures 160–190 ms pixel-to-trajectory latency following the deployment recipe.
Reasoning quality is measured separately: AlpaJudge is a human-calibrated evaluation and reward model that scores reasoning traces, selects training data, and supplies RL rewards. Trajectory metrics describe planning and driving, while reasoning evaluation asks whether an explanation identifies the relevant scene context and supports the decision it describes. On the fixed construction-zone sample, AlpaJudge-RB reports alignment 0.999 and AlpaJudge-RF reports overall 0.842, with five subscores covering action/reference agreement, action safety, component grounding, component completeness, and causal consistency.
Perspective
The demonstrated scope is NVIDIA's latency and paired-quality check on the DRIVE AGX Thor-X reference configuration following the deployment recipe, plus early partners' simulation tests on their own data; the intended reader is a developer preparing to deploy a reasoning driving model in an onboard stack.
The 160–190 ms latency varies with the amount of reasoning rather than being fixed; partner results come from early simulation tests and zero-shot evaluation, and Niulinx's evaluation involved substantial differences in sensor configurations; a high reasoning score only shows that a rationale matches the judge's rubric and must be reported alongside collision, road-departure, and other driving measures.