The Pulse

NVIDIA details 2B-parameter model for onboard AV reasoning

NVIDIA introduced Alpamayo 2 Edge, a vision-language-action model designed for real-time use on its DRIVE AGX Thor platform. The company reports a 160–190 ms pixel-to-trajectory latency for an optimized configuration and describes AlpaJudge

AI.info Team · Written with AI · fact-checked · 2026-11-10

NVIDIA introduced Alpamayo 2 Edge, a 2-billion-parameter vision-language-action model designed for real-time operation on its DRIVE AGX Thor platform. The company says the model combines a Nemotron-based reasoning backbone with a diffusion action expert, using surround-camera observations, temporal context and navigation inputs to generate reasoning and driving trajectories.

It can generate reasoning selectively, NVIDIA says, producing special nominal tokens during routine driving instead of a reasoning trace. The company says closed-loop reinforcement learning in AlpaGym improved driving performance by 48.9% over its best-performing checkpoint without that training.

Model performance and onboard timing

The company’s benchmark figure compares Alpamayo 2 Edge, with 2 billion parameters, against Qwen-Drive-1.0, with 4 billion.

To run the model in real time on DRIVE AGX Thor, the post says it needs further inference optimization. NVIDIA’s deployment recipe uses TensorRT and DFlash2 and documents a reference configuration measured on DRIVE AGX Thor-X with DRIVE OS 7.2.7.0, CUDA 13.4 and TensorRT 11.3.1.9.

The post gives an optimized pixel-to-trajectory latency of 160 to 190 milliseconds, depending on how much reasoning the model generates. The recipe compares optimized outputs with a BF16 PyTorch reference using the same inputs and seeded action noise. Embedl CEO and co-founder Hans Salomonsson said the company got Alpamayo 2 Edge running at 10 Hz on AGX Thor for real-time planning in less than 48 hours.

Scoring model reasoning

NVIDIA describes AlpaJudge as a human-calibrated evaluation and reward model used to assess reasoning traces, choose training data and provide reinforcement-learning rewards. The company offers two variants: AlpaJudge-RB compares a candidate answer with a reference, while AlpaJudge-RF assesses a rationale using camera frames and action context without requiring a reference rationale.

On the notebook’s fixed construction-zone example, NVIDIA reports an RB alignment score of 0.999 and an RF overall score of 0.842. The RF variant also returns five rubric scores covering action agreement, safety, grounding, completeness and causal consistency.

Data selection and training

The post outlines using scores to prioritize examples for another training round. NVIDIA says very low scores can point to a mismatched label, an unsuccessful rollout or a difficult case, while examples in the middle score range may be informative for further training.

The post advises reviewing low-scoring examples before removing them, noting that they can reflect mislabeled data, failed rollouts or difficult cases, including rare failures. It also describes using high scores to find scenes the current model explains consistently.

NVIDIA’s described post-training workflow uses AlpaJudge-RB to score sampled reasoning against reference traces and feed the alignment score into a combined reward alongside trajectory measures. The post says this pattern can add a reasoning signal to open-loop group-relative policy optimization training, while retaining trajectory measures in the reward.