Boston Dynamics taps NVIDIA's Isaac GR00T to power Atlas

For years, Boston Dynamics' Atlas has blown minds with backflips and parkour. But in reality, it behaved more like a world-class gymnast executing choreographed routines than an adaptable worker. To survive the next era of robotics, the world’s most famous hardware needed an equally powerful brain.

That brain is officially arriving. At NVIDIA’s GTC event last week, Boston Dynamics announced an expanded partnership to bring NVIDIA’s Isaac GR00T foundation model framework and Jetson Thor computing platform into the new all-electric Atlas.

Putting Thor inside the electric Atlas

Jetson Thor is NVIDIA’s compact, high-efficiency system designed specifically for the heavy compute needs of humanoids. Boston Dynamics noted that Thor’s power envelope and compact footprint allow Atlas to run complex, multimodal AI models onboard without sacrificing balance or payload.

These multimodal models will interface directly with Boston Dynamics’ proprietary whole-body and manipulation controllers. Boston Dynamics CTO Aaron Saunders called robots the bridge between simulation and the real world, emphasizing that pairing the electric Atlas with Thor provides the machine with the highest-performance compute platform available today.

Beyond hardware, the engineering teams are using Isaac Lab, NVIDIA’s open-source modular framework built on Omniverse and Isaac Sim. Boston Dynamics is using these physically accurate simulated environments to rapidly train AI policies for learned dexterity and locomotion.

Both companies are also working together to establish critical platform benchmarks, including functional safety standards, security architectures, and vision pipelines. Meanwhile, these AI efforts are trickling down to Boston Dynamics' other hardware. Spot has gained reinforcement learning tools for smoother walking and foundation models to dodge specific real-world obstacles.

Why it matters

We think this partnership represents a major reality check and strategic convergence for the entire humanoid industry.

For a long time, purists debated whether hardware giants would build their own end-to-end foundation models or if AI labs would construct their own robot bodies. Boston Dynamics throwing its weight behind GR00T proves that building both frontier hardware and cutting-edge frontier AI under one roof is simply too resource-intensive, even for the industry's pioneer.

It is a massive win for NVIDIA. By securing Atlas as the ultimate poster child for Jetson Thor and Isaac Lab, NVIDIA moves closer to establishing an unshakeable Wintel-like standard for embodied AI. Boston Dynamics, on the other hand, smartly focuses on what it does better than anyone else: mechanical design, actuation, and dynamic physical control.

However, this dependency introduces new questions. Running multimodal foundation models on an untethered humanoid creates severe thermal and battery constraints. While simulation results in Isaac Lab look promising, proving that Atlas can reliably execute autonomous manipulation in unpredictable warehouse environments remains the real test.