NVIDIA upgrades Isaac robot platform, collaborates with Agile Robots to launch humanoid reference platform Agile ONE SN

(By Yang Liang) On October 28, news reported that on October 21, NVIDIA unveiled a major version upgrade of Isaac open robot development platform at GTC Berlin conference, launching a humanoid robot hardware reference platform, new runtime, simulation toolchain, public dataset and open source workflow, connecting the full pipeline from data generation, simulation training to real machine deployment. Jensen Huang, CEO of NVIDIA, stated that the 'ChatGPT moment' in the robotics industry is accelerating.

At the hardware level, Agile Robots (Agile Robots) released the Agile ONE SN humanoid reference platform, deeply adapted to NVIDIA's technology stack, pre-installed with Isaac GR00T robot base model, equipped with Jetson Thor edge computing module, and plans to integrate Halos robot full-stack security system in the future, aimed at universities and research institutions for cutting-edge R&D verification, expected to be officially available in early 2027, with institutions such as Stanford and ETH Zurich expressing interest in using it.

Agile ONE SN physical parameters: height 1.4 meters, total weight 40 kilograms, body with 31 degrees of freedom, single arm maximum load 3 kilograms; equipped with two sets of WUJI Hand 2 five-finger dexterous hands, each hand with 20 degrees of freedom, supporting 1kHz force control and tactile perception. The onboard Jetson Thor can output 2070 FP4 TF computing power, paired with 128GB unified memory, meeting the real-time inference and motion control of the body; the whole machine integrates multi-modal perception devices such as head, wrist cameras, LiDAR, and IMU. The platform integrates simulation, training, and deployment tools, compressing the developer environment setup period from weeks to days.

Software and datasets have seen multiple important updates:

Isaac Lab 3.0 is now officially available, an open-source robot learning framework, supporting multi-GPU distributed training, compatible with Newton and PhysX two physics engines, suitable for robot strategy optimization from desktop to large data centers.

Isaac OS open-source runtime, built on Holoscan, connects AI models, various sensors, motion control, and robot hardware, achieving deterministic low-latency real-time scheduling between simulation environments and physical robots.

SimReady Central resource library is now available on Hugging Face, providing 2000+ factory and robot simulation assets; Human 1M large-scale human dataset, including multi-view synchronized video, 3D hand tracking data, open to full resources for Cosmos alliance partners.

Three sets of DGX Station open-source reference workflows are released, supporting GR00T 1.7 model reinforcement learning fine-tuning, dexterous operation simulation training, and combining ChatGPT Astra for post-training of Cosmos world models, accelerating synthetic robot data production.

Halos full-stack security platform, relying on IGX Thor, provides scalable security capabilities for humanoid, industrial, and warehouse robots, with multiple robot manufacturers already conducting integration evaluations.

A large number of leading robot companies in Europe have joined this ecosystem. ABB and KUKA use Omniverse for factory simulation virtual debugging; Universal Robots and Sereact use GR00T for intelligent goods handling; Hexagon Robotics and Bear Robotics develop localization, mapping, and warehouse operation robots based on Jetson Thor and Isaac OS; Franka Robotics and Pollen Robotics, among other manufacturers, are also using Isaac suite for natural language instruction control and small-scale educational robot projects.

NVIDIA stated that the core challenge in the deployment of general-purpose humanoid robots lies in achieving an adaptive machine that can learn multiple tasks and adapt to human living and working environments. The Isaac full-stack platform is designed to lower the threshold for physical AI development, helping the industry efficiently transfer simulation training results to real hardware, accelerating the industrial deployment process of general-purpose robots.