NVIDIA Releases Alpamayo 2 Super: A 34B Open Vision-Language-Action Model for Robotaxis and Autonomous Driving Under OpenMDW-1.1
NVIDIA released Alpamayo 2 Super, a 34B-parameter vision-language-action (VLA) model for autonomous driving, under the permissive OpenMDW-1.1 license enabling immediate commercial use The architecture combines a 32B Cosmos 3 Super Reasoner VLM backbone with a 2.3B diffusion-based action decoder, trained on ~115,000 hours of multi-camera driving video and over 1 billion images The model produces five outputs in a single forward pass: a planned trajectory (64 waypoints), a Chain-of-Causation (CoC)
Analysis
TL;DR
- NVIDIA released Alpamayo 2 Super, a 34B-parameter vision-language-action (VLA) model for autonomous driving, under the permissive OpenMDW-1.1 license enabling immediate commercial use
- The architecture combines a 32B Cosmos 3 Super Reasoner VLM backbone with a 2.3B diffusion-based action decoder, trained on ~115,000 hours of multi-camera driving video and over 1 billion images
- The model produces five outputs in a single forward pass: a planned trajectory (64 waypoints), a Chain-of-Causation (CoC) trace, a meta-action, reasoning auto-labels, and grounded visual question answering
- It achieves a Lingo-Judge score of 79.2 (ranked first among nearly 40 models), an AlpaSim score of 1.50 ± 0.13, and minADE₆ of 0.911m at 6.4s on challenging benchmarks
- CoC traces integrate with NVIDIA Halos safety-validation workflows and ISO/PAS 8800 compliance, while the model can serve as an autolabeler that compresses annotation cycles from months to days
Why It Matters
Alpamayo 2 Super represents a significant step toward unified VLA models for autonomous driving that can simultaneously reason, plan, and explain decisions — addressing the long-tail event problem that plagues conventional detection-and-prediction stacks. Its open commercial licensing and demonstrated ability to generate causal explanations make it directly relevant to teams building safe, auditable autonomous systems. The autolabeling capability also offers immediate operational value by dramatically reducing data annotation costs for fleet operators.
Technical Details
- Architecture: 34B total parameters split into a 32B VLM backbone (built on NVIDIA Cosmos 3 Super Reasoner, post-trained with reinforcement learning) and a 2.3B diffusion-based action decoder. Tested on a single H100 80GB at 72,115 MiB peak memory; distillation recommended for in-car deployment.
- Inputs and Outputs: Inputs include multi-camera RGB video (validated with six cameras, four historical frames per camera), text, and egomotion history (3D translation + 3×3 rotation matrix, multi-timestep). The trajectory API outputs 64 waypoints spanning 0.1–6.4 seconds at 0.1s intervals, each with ego-frame XYZ and a 3×3 rotation matrix.
- Training Data: Approximately 115,000 hours of multi-camera driving video with egomotion and trajectory annotations, ~3.7 million Chain-of-Causation (CoC) traces providing structured causal explanations of driving decisions, and over one billion training images.
- Benchmarks: LingoQA Lingo-Judge score of 79.2 (first among ~40 models, outperforming Qwen2.5-VL 72B by 17.0, Gemini 2.5 Pro by 15.1, and GPT-4o by 23.2). Closed-loop AlpaSim score of 1.50 ± 0.13 on 910 PhysicalAI-AV-NuRec scenarios. Open-loop minADE₆ of 0.911m at 6.4s on 937 challenging PhysicalAI-AV samples.
- Five-Output Design: A single forward pass produces (1) planned trajectory, (2) CoC causal explanation trace, (3) meta-action (e.g., yield, lane change), (4) reasoning auto-labels, and (5) visual question answering with 2D grounding. CoC traces integrate with NVIDIA Halos for safety validation and ISO/PAS 8800 alignment.
Industry Insight
- The open commercial licensing (OpenMDW-1.1 + Apache 2.0) signals NVIDIA's strategy to accelerate ecosystem adoption of its VLA stack, enabling competitors and partners to build derivative models and commercial products without additional permission — a move that could standardize open VLA architectures in autonomous driving.
- The integration of Chain-of-Causation traces with safety-validation workflows addresses a critical industry bottleneck: regulatory compliance and explainability. Teams prioritizing ISO/PAS 8800 alignment should evaluate Alpamayo 2 Super as both a planning model and a safety audit tool.
- The autolabeling claim (compressing annotation cycles from months to days) represents a compelling ROI argument for fleet operators with proprietary data; even conservative estimates of annotation cost savings could justify early adoption and fine-tuning on domain-specific driving data.
Disclaimer: The above content is generated by AI and is for reference only.