National Robotics Week - Latest Physical AI Research, Breakthroughs and Resources

Check back here throughout the week to learn the latest on physical AI, which enables machines to perceive, plan and act with greater autonomy and intelligence in real-world environments.

This National Robotics Week, running through April 12, NVIDIA is highlighting the pioneering technologies that are shaping the future of intelligent machines and driving progress across manufacturing, healthcare, logistics and more.

Advancements in robotics simulation and robot learning are driving this fundamental shift in the industry. Plus, the emergence of world foundation models is accelerating the evolution of AI-enabled robots capable of adapting to dynamic and complex scenarios.

For example, by providing robot foundation models like NVIDIA GR00T N1, frameworks such as NVIDIA Isaac Sim and Isaac Lab for robot simulation and training, and synthetic data generation pipelines to help train robots for diverse tasks, the NVIDIA Isaac and GR00T platforms are empowering researchers and developers to push the boundaries of robotics.

Spotlight on Wheeled Lab: Advancing Simulation-to-Reality Robotics With NVIDIA Isaac Lab Wheeled Lab, a research project from the University of Washington, is bringing simulation-to-reality robotics to low-cost, open-source platforms.

Wheeled Lab, integrated with NVIDIA Isaac Lab - a unified framework for robot learning - enables reinforcement learning models to train wheeled robots for complex tasks like controlled drifting, obstacle avoidance, elevation traversal and visual navigation. This pipeline uses domain randomization, sensor simulation and end-to-end learning to bridge the gap between simulated training and real-world deployment, all while ensuring zero-shot simulation-to-reality transfer.

The entire stack - spanning simulation, training and deployment - is fully open source, giving developers the freedom to iterate, modify policies and experiment with reinforcement learning techniques in a reproducible environment.

Get started with the code on GitHub.

Teaching Robots to Think: Nicklas Hansen's AI Breakthroughs What does it take to teach robots complex decision-making in the real world? For Nicklas Hansen, a doctoral candidate at UC San Diego and an NVIDIA Graduate Research Fellow, the answer lies in scalable, robust machine learning algorithms.

With experience from the University of California, Berkeley, Meta AI (FAIR) and the Technical University of Denmark, Hansen is pushing the boundaries of how robots perceive, plan and act in dynamic environments. Their research sits at the intersection of robotics, reinforcement learning and computer vision - bridging the gap between simulation and real-world deployment.

Nicklas Hansen, a doctoral candidate at UC San Diego and an NVIDIA Graduate Research Fellow. Hansen's recent work tackles one of robotics' toughest challenges: long-horizon manipulation. Their paper, Multi-Stage Manipulation With Demonstration-Augmented Reward, Policy and World Model Learning, introduces a framework that enhances data efficiency in sparse-reward environments by using multistage task structures.

Left: A simulated Franka robot solves a peg insertion manipulation task. Right: Hansen's method, DEMO3, infers task progress directly from raw visual observations. Another key project of Hansen's, Hierarchical World Models as Visual Whole-Body Humanoid Controllers, advances control strategies for humanoid robots, enabling more adaptive and humanlike movements.

Beyond their own research, Hansen advocates for making AI-driven robotics more accessible.

My advice to anyone looking to get started with AI for robotics is to simply play around with the many open-source tools available and gradually start contributing to projects that align with your goals and interests, they said. With the availability of free simulation tools like MuJoCo, NVIDIA Isaac Lab and ManiSkill, you can make a profound impact on the field without owning a real robot.

Hansen is the lead author of TD-MPC2, a model-based reinforcement learning algorithm capable of learning a variety of control tasks without any domain knowledge. The algorithm is open source and can be run on a single consumer-grade GPU.

Learn more about Hansen and other NVIDIA Graduate Fellowship recipients driving innovation in AI and robotics. Watch a replay of the Graduate Program Fast Forward session from the NVIDIA GTC AI conference, where doctoral students in the NVIDIA Graduate Fellowship showcased their groundbreaking research.

Hackathon Features Robots Powered by NVIDIA Isaac GR00T N1 The Seeed Studio Embodied AI Hackathon, which took place last month, brought together the robotics community to showcase innovative projects using the LeRobot SO-100ARM motor kit.

The event highlighted how robot learning is advancing AI-driven robotics, with teams successfully integrating the NVIDIA Isaac GR00T N1 model to speed humanoid robot development. A notable project involved developing leader-follower robot pairs capable of learning pick-and-place tasks by post-training robot foundation models on real-world demonstration data.

How the project worked:

Real-World Imitation Learning: Robots observe and mimic human-led demonstrations, recorded through Arducam vision systems and an external camera.

Post-Training Pipeline: Captured data is structured into a modality.json dataset for efficient GPU-based training with GR00T N1.

Bimanual Manipulation: The model is optimized for controlling two robotic arms simultaneously, enhancing cooperative skills.

The dataset is now publicly available on Hugging Face, with implementation details on GitHub.

Team Firebreathing Rubber Duckies celebrating with NVIDIA hosts. Learn more about the project.

Advancing Robotics: IEEE Robotics and Automation Society Honors Emerging Innovators The IEEE Robotics and Automation Society in March an

LINK:	https://blogs.nvidia.com/blog/national-robotics-week-2025/...
	See more stories from nvidia

North America Stories