Google & NVIDIA both shipped open source robot stacks at ROSCon. Only one supports humanoids

Free software is good

Google and NVIDIA are both giving the robotics community free and open source software for robots.

Intrinsic Core, Google's release, is a set of ROS-compatible capabilities under Apache 2.0 that includes a real-time control framework, motion planning, grasp planning, 6-DoF pose estimation, Gazebo-backed simulation, camera calibration, a native digital twin and pre-configured drivers. It ships with a reference design for CNC machine tending, with support for FANUC and Universal Robots hardware.

It is also, as The Robot Report noted, "a platform for robotic arms; it doesn't support mobile robots or humanoids."

NVIDIA's Isaac ROS 5.0, announced the same day and also free and open source, lists Mentee Robotics' MenteeBot humanoid among its deployments. It's a set of GPU-accelerated packages that sit on top of ROS that handle the compute-heavy parts of a robot's perception and motion, including object pose detection and tracking, stereo depth, spatial mapping and collision-free arm path planning. Interestingly, it's agentic ... so AI coding agents, not just human engineers, can wire up perception and manipulation pipelines.

What NVIDIA shipped

Isaac ROS 5.0 is the more technically interesting of the two releases, and the more relevant to humanoid robots.

The headline addition is ROS Lyrical and Ubuntu 24.04 support. NVIDIA contributed a standard data-handling interface to ROS Lyrical itself, developed with the Open Source Robotics Alliance, with CUDA as the working example of GPU acceleration through it. 

That's NVIDIA putting plumbing into ROS core rather than shipping packages that sit on top of it.

On the perception side, FoundationPose now ships as an agent-ready inference library and tracks object position and orientation up to 5.5X faster than the previous release. FoundationStereo gets a fine-tuning skill that adapts stereo perception models to a developer's specific cameras and environment, which is the unglamorous problem that eats weeks on every new robot build. cuMotion, NVIDIA's GPU-accelerated motion planner, maps collision-free paths in roughly 2 to 5 milliseconds in the implementation robotics software engineering firm Ekumen demonstrated.

The word NVIDIA keeps using is "agentic." 

Isaac ROS 5.0 ships agent-ready documentation, so AI coding agents can parse the toolkit's workflows. In addition, there are reusable skills that a human or an agent can invoke to do development work: a standalone pick-and-place skill that chains detection, depth estimation and pose output, a setup skill, the FoundationStereo fine-tuning skill. There's also an integration with AgenticROS, a RealSense-sponsored project connecting Isaac ROS to NVIDIA's Nemotron open models and NemoClaw blueprints so agents can operate ROS-based robots directly.

Intrinsic Core vs ROS 5.0

In plain terms: Intrinsic open sourced better building blocks for human robotics engineers. NVIDIA open sourced building blocks plus the machine-readable scaffolding for AI agents to assemble them. 

Given that both companies are pitching a developer shortage as the reason robotics adoption is slow, those are meaningfully different answers to related problems.

What humanoid teams can and can't take from it

Mentee Robotics uses Isaac ROS as the perception and AI backbone of MenteeBot, interpreting visual information and executing learned behaviors in real time, across both Jetson Orin and Jetson Thor.

Isaac ROS gives a humanoid team perception, object pose, stereo depth, spatial mapping and GPU-accelerated arm motion planning. It does not give them balance, gait, whole-body control or fall recovery. cuMotion is arm motion planning. 

Nobody shipped a bipedal locomotion stack at ROSCon this week, but there is now an open infrastructure layer that covers a humanoid's eyes and arms.

Just not the legs.

Where Google's humanoid work lives instead

Google has a humanoid strategy, of course. 

It just doesn't run through Intrinsic, and it isn't open.

Gemini Robotics 2, which Google DeepMind shipped on July 30, covers whole-body humanoid control, embodied reasoning and on-device execution, with Apptronik, Boston Dynamics, Agile Robots and Franka among launch partners. Access is by application for the VLA and on-device models, private preview plus AI Studio for the reasoning model, trusted-tester enrollment for the SDK. There are no open weights.

Which leaves Google running two robotics programs with inverted openness: the arm stack is free to anyone, the whole-body stack is invitation-only. 

I wrote about what that split buys Google at Forbes.

Worth noting the two companies aren't really opponents at this level. Intrinsic Core's pose estimation runs on NVIDIA FoundationPose, and Intrinsic shows up on NVIDIA's own partner list for Isaac ROS 5.0, cited for machine tending. 

What differs is reach. NVIDIA builds horizontally across every robot morphology it can get a Jetson into, because it sells the compute under all of them. Intrinsic builds vertically into manipulation-based manufacturing, because that's where its revenue is.

Why the legs stay closed

Six-axis arm kinematics is a solved problem with decades of literature, stable hardware interfaces and customers who will pay today. Giving away a motion planner for a FANUC arm costs Intrinsic nothing it could have monetized and potentially buys it an ecosystem. 

Perception is similar: NVIDIA wants every robot seeing through CUDA, and the more that happens ... the more Jetsons ship.

Whole-body control for a biped is neither. It's largely unsolved, at least at an open source level, and it's tightly coupled to specific actuators and mass distribution. For most humanoid companies it remains a real share of what the valuation rests on. Figure, 1X, Agility, Apptronik and Unitree have not open sourced their balance and locomotion stacks, and there's no obvious commercial reason they would.

So every humanoid company keeps building that layer in-house, largely from scratch, largely in parallel, solving overlapping problems. 

That's a duplicated cost the arm world is now at least partially shedding.

The counterargument is that it's early and it's supposed to work this way. Redundant effort at the infrastructure layer is the normal cost of a sector where infrastructure is still the product. The arm world only got a shared foundation once the fundamentals stopped being competitive.

However, the day whole-body control becomes commodity plumbing instead of proprietary advantage is the day this industry's cost structure changes. NVIDIA, with GR00T weights already public, a humanoid on the Isaac ROS 5.0 list and roughly 1.3 million ROS users to distribute to, is the most plausible party to push it there.

We will, as they say, see.