NVIDIA unveils new physical‑AI stack
NVIDIA announced a raft of robotics/embodied AI tools at GTC — Isaac GR00T N1.7 and Alpamayo 1.5 VLA models, the Cosmos 3 world model, plus Newton 1.0 GA for contact‑rich sim-to-real learning — all aimed at scaling perception-to-action on real robots reported and documented by NVIDIA. The pitch: swap robotics' data bottleneck for compute and unified world models so robots generalize faster in factories and fleets.
Newton 1.0 GA announced) at NVIDIA GTC on March 16, 2026 and ships as an open‑source, GPU‑accelerated physics engine built on NVIDIA Warp and OpenUSD for contact‑rich manipulation and locomotion. Newton’s MuJoCo Warp and other optimizations report throughput speedups of ~252× for locomotion and ~475× for manipulation compared with MJX on an NVIDIA RTX PRO 6000 Blackwell, according to NVIDIA’s technical blog. (developer.nvidia.com) GR00T N1.7 was released in early access with commercial licensing to support production robot deployments, per NVIDIA’s GTC press materials dated March 16, 2026. (nvidianews.nvidia.com) At the same keynote NVIDIA previewed a next‑gen GR00T N2 research path claimed to outperform prior VLA models on novel tasks (previewed performance claims reported in coverage of the keynote). (the-decoder.com) Alpamayo 1.5 is a 10B‑parameter chain‑of‑thought vision‑language‑action model that was updated with a Cosmos‑Reason2 backbone, RL post‑training, text‑guided trajectory planning and flexible multi‑camera support, per the Hugging Face community post. (huggingface.co) NVIDIA’s Alpamayo repo and docs note the model weights are substantial (example download ~22 GB) and recommend GPUs with ≥24 GB VRAM (e.g., A5000, H100, RTX 4090) for inference and fine‑tuning. (github.com) Cosmos 3 is described in NVIDIA’s GTC materials as the first “world foundation model” to unify controllable synthetic world generation, physical reasoning and action simulation for physical AI. (investor.nvidia.com) Early Cosmos adopters named by NVIDIA include Agility Robotics, Figure AI, Foretellix, Skild AI and Uber, which NVIDIA says are using Cosmos and Omniverse blueprints to generate scalable, physics‑grounded synthetic data for robot and AV training. (investor.nvidia.com) NVIDIA positioned these releases as an integrated stack: DGX/Omniverse/Cosmos for large‑scale model training and synthetic data, Newton/Isaac Sim for physics validation, and Drive/Alpamayo for vehicle deployment, with the GTC program running March 16–19, 2026. (nvidianews.nvidia.com) The Cosmos cookbook and GitHub repos provide post‑training scripts and APIs to connect WFMs to Isaac simulation pipelines for post‑training and closed‑loop validation. (github.com) NVIDIA’s coverage lists ecosystem partnerships and integration targets—industrial robot OEMs and humanoid builders (AGIBOT, Agility, Figure, FANUC, ABB, KUKA, Skild) plus automakers and mobility providers standardizing on DRIVE Hyperion for Level‑4 stacks—signaling coordinated toolchain adoption rather than single‑product rollouts. (tmcnet.com)
