NVIDIA ships Alpamayo 2 Edge, a 2B reasoning VLA model for in-vehicle AV deployment
huggingface · x · 2026-11-12
NVIDIA released Alpamayo 2 Edge, a 2B-parameter adaptive reasoning vision-language-action model built for real-time deployment on DRIVE AGX Thor.
Key points:
- Generates Chain-of-Causation reasoning on demand and uses a diffusion action expert to produce driving trajectories from surround cameras and navigation inputs
- Closed-loop RL in AlpaGym improves driving performance by 48.9% over the best non-RL checkpoint
- AlpaJudge offers human-calibrated reasoning evaluation, usable for selecting training examples and as post-training rewards
- TensorRT + DFlash2 optimization achieves 160-190 ms pixel-to-trajectory latency on DRIVE AGX Thor
The model and AlpaJudge are available on Hugging Face.
Related event: NVIDIA Open-Sources Alpamayo 2 Edge, a 2B-Parameter In-Car VLA Model(7 posts)→
More from Embodied
- OSU open-sources FAME-QP, a GPU QP solver cutting MPC latency for humanoid and quadrotor control — ohiostate · 2026-11-12
- Google, Samsung, Gentle Monster smart glasses open pre-orders, ship Dec 10 — realkaranahuja · 2026-11-12
- The Verge hands-on: Apple's $299 HomeView smart display, tested with Siri AI — The Verge AI · 2026-11-12
- Fruit Fly Brain Connectome of 166,000 Neurons Translated Into Robot Control — CurieuxExplorer · 2026-11-12
- CoRL 2026 SPIN workshop talk: robots learning through interaction and why structure aids scale — GeorgiaChal · 2026-11-12
- Fable 5.2 digital twin nails real-robot rollout, aiming for sim RL training and evals — DominiqueCAPaul · 2026-11-12