NVIDIA unveiled a comprehensive robotics development platform at the Conference on Robot Learning (CoRL) in Seoul, introducing open-source tools designed to accelerate the transition from research labs to real-world robotic applications. The semiconductor giant’s latest announcements center on solving fundamental challenges in robotics development: creating realistic simulations for training, enabling human-like reasoning in robots, and bridging the gap between virtual testing and physical deployment.
The new platform combines three core technologies that work together as an integrated system. The Newton Physics Engine provides realistic simulation environments where robots can safely learn complex tasks. The Isaac GR00T foundation model serves as the “brain” that enables robots to understand instructions and reason through problems. Meanwhile, Cosmos world foundation models generate diverse training data to help robots adapt to countless real-world scenarios.
Robotics developers have long struggled with a critical bottleneck: teaching robots skills in the real world is expensive, time-consuming, and potentially dangerous. Simulation offers a solution, but existing physics engines often fail to accurately represent the complex dynamics of humanoid robots navigating real environments.
NVIDIA’s Newton Physics Engine, developed in partnership with Google DeepMind and Disney Research, addresses this challenge by providing GPU-accelerated simulations that can model intricate physical interactions. The open-source engine, now managed by the Linux Foundation, enables developers to simulate complex scenarios like robots walking through snow or carefully handling delicate objects—skills that can then transfer reliably to physical robots.
The engine’s significance becomes clear when considering that over 250,000 robotics developers worldwide need accurate physics simulation for their work. Newton’s flexible architecture supports different types of physics solvers, allowing researchers to model everything from basic locomotion to sophisticated manipulation tasks.
Early adopters include prestigious institutions like ETH Zurich’s Robotic Systems Lab, Technical University of Munich, and Peking University, alongside companies such as robotics firm Lightwheel and simulation specialist Style3D. These organizations are integrating Newton into their development workflows to accelerate robot training while reducing the risks associated with physical testing.
Traditional robots excel at repetitive, programmed tasks but struggle with the ambiguous instructions and unexpected situations that define real-world environments. NVIDIA’s Isaac GR00T N1.6 foundation model tackles this limitation by incorporating reasoning capabilities that mirror human problem-solving approaches.
Foundation models represent a significant advancement over conventional AI systems. Rather than being trained for specific tasks, these models learn broad patterns from massive datasets, enabling them to generalize across diverse situations. For robotics, this means a single model can potentially handle multiple types of tasks rather than requiring separate programming for each function.
The latest Isaac GR00T model integrates NVIDIA’s Cosmos Reason technology, which has been downloaded over one million times and currently leads Hugging Face’s Physical Reasoning Leaderboard. This integration allows robots to break down vague instructions into step-by-step plans, using common sense and physics understanding to navigate unfamiliar scenarios.
For example, when instructed to “tidy up the living room,” the model can interpret this broad command by identifying objects that are out of place, determining appropriate locations for each item, and planning efficient movement patterns—all while adapting to obstacles or unexpected conditions it encounters.
The N1.6 version introduces enhanced capabilities for simultaneous movement and object manipulation, giving humanoid robots greater flexibility in their torso and arm movements. This improvement enables more complex tasks like opening heavy doors while maintaining balance and spatial awareness.
Major robotics companies are already evaluating Isaac GR00T for their platforms. Partners include AeiROBOT, Franka Robotics, LG Electronics, Lightwheel, Mentee Robotics, Neura Robotics, Solomon, Techman Robot, and UCR—representing a broad cross-section of the commercial robotics industry.
One of the biggest challenges in AI development is obtaining sufficient high-quality training data. For robotics, this problem is particularly acute because robots must learn to operate in countless environmental variations that would be impractical to capture through real-world data collection alone.
NVIDIA’s Cosmos world foundation models address this challenge by generating synthetic training data from text, image, and video prompts. These models have been downloaded over three million times, indicating strong developer adoption across the robotics community.
The upcoming Cosmos Predict 2.5 combines three separate Cosmos models into a single, more efficient system. This consolidation reduces complexity while expanding capabilities, including the ability to generate up to 30-second videos and multi-camera perspectives for more comprehensive training scenarios.
Cosmos Transfer 2.5 focuses on creating photorealistic synthetic data from 3D simulation scenes. Despite being 3.5 times smaller than previous models, it delivers faster, higher-quality results while supporting various spatial control inputs like depth mapping, segmentation, and high-definition environmental maps.
These capabilities enable developers to create vast libraries of training scenarios without the expense and logistical challenges of real-world data collection. A robot learning to navigate indoor environments, for instance, can train on thousands of synthetic home layouts, lighting conditions, and obstacle configurations before ever encountering a physical space.
Object manipulation represents one of robotics’ most persistent challenges. Teaching a robot to grasp objects reliably requires understanding complex physics interactions, adapting to different object properties, and maintaining control through dynamic movements.
NVIDIA’s new dexterous grasping workflow, available in Isaac Lab 2.3 developer preview, uses an automated curriculum approach to teach multi-fingered robotic hands. The system begins with simple grasping tasks and gradually introduces complexity by varying factors like gravity, friction, and object weight.
This progressive learning approach mirrors how humans develop motor skills, starting with basic movements and building toward more sophisticated manipulation abilities. The virtual training environment can simulate countless object types and environmental conditions, enabling robots to develop robust grasping strategies.
Boston Dynamics has already implemented this workflow to enhance its Atlas robot’s manipulation capabilities, demonstrating the practical value of the approach. Other major robotics companies adopting NVIDIA’s Isaac and Omniverse technologies include Agility Robotics, Figure AI, Hexagon, Skild AI, Solomon, and Techman Robot.
Validating robot performance presents another significant challenge for developers. Testing skills on physical robots is slow and expensive, while simplified simulation environments often fail to predict real-world performance accurately.
NVIDIA is collaborating with Lightwheel to develop Isaac Lab-Arena, an open-source policy evaluation framework designed for large-scale testing in realistic simulated environments. This framework will enable developers to run comprehensive evaluations across multiple scenarios and tasks without building custom testing infrastructure.
The framework addresses a common problem in robotics development: fragmented, oversimplified testing that doesn’t reflect real-world complexity. By providing standardized, comprehensive evaluation tools, Isaac Lab-Arena should help bridge the gap between simulation performance and real-world deployment success.
NVIDIA’s robotics platform requires substantial computational resources, prompting the company to announce new AI infrastructure specifically designed for robotics development workloads.
The GB200 NVL72 represents NVIDIA’s most powerful training system, integrating 36 Grace CPUs with 72 Blackwell GPUs in a rack-scale configuration. Major cloud providers are adopting this system to support complex AI training tasks, including the reasoning and physical AI applications central to modern robotics.
For development teams requiring on-premises capabilities, NVIDIA RTX PRO Servers offer a unified architecture supporting the full robotics development pipeline, from training and synthetic data generation to simulation and testing. The RAI Institute has already adopted these servers for their robotics research initiatives.
At the edge, Jetson Thor brings Blackwell GPU capabilities to robots themselves, enabling real-time on-robot inference for multi-AI workflows. This capability is crucial for humanoid robots that must process visual, auditory, and sensor data simultaneously while making split-second decisions. Partners implementing Jetson Thor include Figure AI, Galbot, Google DeepMind, Mentee Robotics, Meta, Skild AI, and Unitree.
The academic research community’s adoption of NVIDIA’s robotics technologies demonstrates their scientific credibility and practical value. Nearly half of the papers accepted at CoRL reference NVIDIA technologies, including GPUs, simulation frameworks, and CUDA-accelerated libraries.
Notable research projects highlighted at the conference include BEHAVIOR, a robotic learning benchmark developed by Stanford’s Vision and Learning Lab, and Taccel, a high-performance simulation platform for vision-based tactile robotics created by Peking University. These projects represent cutting-edge research that will likely influence commercial robotics development in coming years.
The broad academic adoption across institutions like Carnegie Mellon, University of Washington, ETH Zurich, and the National University of Singapore suggests that NVIDIA’s approach aligns with the research community’s understanding of robotics development challenges and solutions.
NVIDIA’s comprehensive robotics platform represents a significant shift toward standardized, open-source development tools for the robotics industry. By providing integrated simulation, reasoning, and evaluation capabilities, the platform could accelerate the transition from research prototypes to commercial applications.
The open-source approach is particularly noteworthy in an industry where proprietary solutions have traditionally dominated. This strategy could foster broader innovation by enabling smaller companies and research institutions to access sophisticated development tools previously available only to well-funded organizations.
For businesses considering robotics investments, NVIDIA’s platform offers a potentially more predictable development path. The integrated nature of the tools should reduce the technical risks associated with combining disparate technologies from multiple vendors, while the strong academic adoption suggests robust long-term support and continued innovation.
The emphasis on humanoid robots also signals NVIDIA’s belief that general-purpose robotic assistants represent the next major market opportunity, moving beyond the industrial automation applications that have dominated robotics to date.
Join hosts Anthony, Shane, and Francesca for essential insights on AI's impact on jobs, careers, and business. Stay ahead of the curve – listen now!
Join hosts Anthony, Shane, and Francesca for essential insights on AI's impact on jobs, careers, and business. Stay ahead of the curve – listen now!
No hype. No doom. Just actionable resources and strategies to accelerate your success in the age of AI.
