Developers can now train, simulate and deploy robots that perceive, reason and act across multiple tasks in industrial and service applications.
As robotics expands across industries, developers face growing challenges in building AI-powered machines that can perceive, reason and act in the physical world. Conventional robots are often limited to single tasks, require extensive programming and depend on costly resources for training and evaluation, slowing innovation and deployment.
To address this, NVIDIA unveiled new physical AI models, frameworks and infrastructure, alongside robots from global partners including Boston Dynamics, Caterpillar, Franka Robotics, Humanoid, LG Electronics and NEURA Robotics. The announcements focus on accelerating the development of “generalist-specialist” robots capable of quickly learning multiple tasks, from industrial manipulation to household assistance.
NVIDIA introduced open models including Cosmos Transfer 2.5 and Cosmos Predict 2.5 for synthetic data generation and policy evaluation, Cosmos Reason 2 for reasoning and vision-language understanding, and Isaac GR00T N1.6 for humanoid full-body control. These models are integrated into the Hugging Face LeRobot library, enabling developers to fine-tune robots efficiently. Scalable simulation and training are supported through Isaac Lab-Arena, an open-source evaluation framework and OSMO, a cloud-native orchestration platform for robot workflows.
The Jetson T4000 module, powered by NVIDIA’s Blackwell architecture, delivers four times the efficiency and AI compute of its predecessor, while IGX Thor extends robotics AI to industrial edge environments. Partners are already using these technologies to build advanced robots, from NEURA’s dexterous humanoids to LEM Surgical’s AI-guided surgical arms and Caterpillar’s autonomous construction equipment.
Key products and their features include:
- Open physical AI models for robot learning, reasoning and action
- Isaac Lab-Arena for large-scale simulation and benchmarking
- OSMO cloud-native orchestration for streamlined workflows
- Jetson T4000 and Thor modules for high-efficiency, high-performance AI computing
- Integration with Hugging Face LeRobot library for rapid open-source development
Jensen Huang, NVIDIA founder, says, “The ChatGPT moment for robotics is here. Breakthroughs in physical AI, models that understand the real world, reason and plan actions are unlocking entirely new applications.”
