The Robot Race Left the Hardware: NVIDIA, Google and Qualcomm All Just Open-Sourced the Same Fight
Introduction
For most of the last decade, the argument about who is winning in robotics has been settled by counting things: how many arms are bolted to a line, how many humanoids rolled off a conveyor, how many autonomous vehicles are licensed. Hardware was the scoreboard. At the ROSCon developer conference in Toronto in late September, three of the largest names in physical AI made the same move within roughly 48 hours — and none of them were about hardware.
On September 22, NVIDIA released Isaac ROS 5.0, a collection of GPU-accelerated packages built on the Robot Operating System, and framed a substantial part of the release around AI agents writing robot software. On the same day, Intrinsic — the robotics software unit of Alphabet, folded into Google earlier in the year — published the core of its industrial stack under a permissive Apache 2.0 licence. On September 23, Qualcomm announced an agreement to acquire PickNik Inc., the Boulder, Colorado company that has stewarded MoveIt, one of the most widely used open-source motion-planning frameworks in existence.
Three companies. Three different strategies — accelerate it, donate it, buy it — and one identical conclusion: the contested layer in physical AI is now software, and the incumbents have decided the fastest way to win it is to give it away. That is a genuinely new phase for the industry, and it follows a run of hardware-first announcements that [our recent coverage of the robotics and drones beat] has tracked closely.
Main Content
NVIDIA ships the agents
NVIDIA's release is the most interesting of the three because it changes who writes the code. Isaac ROS 5.0 is not primarily a new perception model; it is a set of agentic development workflows. Reusable skills handle the tedious parts of a robotics project — setting up an environment, pick-and-place sequences, adapting a perception model to a specific camera. The stated effect is that an AI agent, rather than a human engineer, can carry out those steps, supported by documentation written specifically to be legible to an agent.
The measurable claim attached to it is specific: FoundationPose, the pose-estimation model NVIDIA has pushed for six-degree-of-freedom object tracking, now detects and tracks the position and orientation of objects up to 5.5 times faster than before. The release adds support for ROS Lyrical, the current flagship distribution, and for Ubuntu 24.04. It runs across the Jetson line from the Orin Nano up to Thor. Critically, it remains free and open source, and NVIDIA says it is available to what it describes as nearly 1.3 million ROS users. Universal Robots, Magna and Intrinsic are named as organisations already building on the software.
One detail deserves more attention than it got. NVIDIA says it worked with the Open Source Robotics Alliance to contribute a standard data-handling interface to ROS Lyrical, so that robotics software runs efficiently across different compute hardware including GPUs. That is a standards-body contribution rather than a product feature, and it is arguably the more consequential of the two — it means the performance advantage of a proprietary silicon stack is being papered over at the interface layer, which is exactly what a platform war looks like when the challenger has more engineers than the incumbent.
Intrinsic gives away the factory
Intrinsic Core is the opposite bet: not an acceleration story but a completeness story. Alphabet acquired Intrinsic in 2021 and folded it into Google in February, and the unit's stack had been used in real manufacturing deployments but was not available to outside developers. Intrinsic Core is a set of ROS-compatible capabilities drawn from that production stack, published under Apache 2.0 on GitHub.
What is actually in the package is worth listing, because it describes what a complete industrial robotics stack has to contain. There is a hardware-agnostic real-time control framework, asset models and a native digital twin, pose detection, motion and grasp planning, Gazebo-based simulation, camera calibration, and ROS drivers. Alongside it is an open reference solution for feeding CNC machines with robots built by FANUC and Universal Robots.
Alongside the core, Intrinsic also published the Open Machine Tending Solution, a ready-made reference design for a machine shop that runs on both Intrinsic Core and the Open Robotics Suite. The argument Intrinsic makes is a market-sizing one: it says its first AI for Industry Challenge drew more than 5,000 developers and roboticists across 115 countries, and that in the US and Europe alone there are tens of thousands of businesses fabricating parts of which only a small fraction have any automation at all.
The quote that best explains the strategy comes from Brian Gerkey, Intrinsic's chief technology officer, describing the release as "the core of our own software stack." That is an unusually candid framing. The gift is also a demonstration — if a prospective customer's first experience is the same code the company runs in its own plants, the sales conversation changes.
There is a limit worth stating plainly. Intrinsic Core does not yet cover mobile robots or humanoids. It is a manipulation and factory-automation release, not a general physical-AI platform, and the humanoid programmes that dominate this year's headlines are running on somebody else's stack.
Qualcomm buys the planner
The third move is the most aggressive. Qualcomm announced an agreement to acquire PickNik, the company behind MoveIt. MoveIt is motion-planning software for robotic arms — the layer that decides how a jointed arm actually reaches a pose without colliding with anything — and it is close to universal in academic, industrial and commercial robotics. It has two open-source releases, MoveIt 1 and MoveIt 2, plus a commercial tier called MoveIt Pro.
Qualcomm did not disclose the purchase price, and the transaction remains subject to customary closing conditions. What it did commit to publicly is that MoveIt 1 and MoveIt 2 remain open, community-driven and supported across third-party hardware. The commercial product, MoveIt Pro, becomes more tightly integrated with Dragonwing, Qualcomm's robotics platform, beginning with the Arduino VENTUNO Q boards.
Nakul Duggal, Qualcomm's executive vice president and group general manager for automotive, industrial and embedded IoT and robotics, framed it in the language of connective tissue: "Robotics is one of the most exciting frontiers of physical AI, and software is the connective tissue that turns great hardware into great robots." PickNik's founder and chief product officer, Dave Coleman, pointed at the part of the business that was actually bought: investment in MoveIt as a training platform for robotics AI models. Its CEO, Dave Grant, said only that MoveIt is "entering an exciting new chapter with the resources and support to help expand its impact."
The full release is on Qualcomm's own newsroom, and the reasoning behind the open-sourcing commitment is set out in Intrinsic's announcement of Intrinsic Core for the parallel strategy at Google. Both Qualcomm and PickNik are founding members of the Open Source Robotics Alliance, and Qualcomm says it intends to keep backing Space ROS, a hardened build of ROS aimed at critical applications including space exploration — which tells you how seriously the company intends to be taken as an upstream contributor rather than a downstream consumer.
What it means for integrators
Put the three announcements side by side and a coherent strategy emerges that nobody was running six months ago. NVIDIA supplies the acceleration layer and wants to own the developer relationship through agentic tooling. Qualcomm supplies the silicon and buys the planning layer outright, keeping it free so that Dragonwing becomes the default choice for anyone starting a new arm project. Intrinsic supplies the complete application layer and donates it because the market it needs to win is the long tail of job shops that cannot afford an integrator.
Each of them is giving away something in order to make the paid thing underneath more default. That is what a standards war looks like from inside, and it is a reason for integrators to be cautiously optimistic rather than nervous. Every one of the three commitments — Isaac ROS under open source, Intrinsic Core under Apache 2.0, MoveIt 1 and 2 remaining open — is a promise, not a contract, and all three companies have commercial reasons to revisit it. But the direction of travel is unambiguous, and it favours the small and mid-sized integrator who has been priced out of the robotics stack until now.
The direction of travel was already visible at Hannover Messe earlier this year, where Qualcomm showed its IQ10 processor series for autonomous mobile robots and humanoids; the HMND 01 humanoid from startup Humanoid completed a logistics pilot at Siemens' Erlangen plant running on Jetson Thor; SCHUNK presented an automation cell simulated with Omniverse and Isaac; Hexagon built its AEON humanoid on NVIDIA technology; and Dresden's Wandelbots showed a NOVA platform pursuing something close to Intrinsic's approach with a unified control layer across robots from different manufacturers. ROSCon did not start that convergence. It formalised it. The trade-press roundup of the Toronto announcements makes the same read from the industrial exhibition side of the fence.
Conclusion
Robotics has spent a decade competing on metal and compute. In three announcements made inside 48 hours at a developer conference in Toronto, NVIDIA, Google and Qualcomm each conceded that the durable advantage in physical AI lives one layer higher — in the software that developers actually touch — and each chose to compete by giving that layer away.
There is a real risk in that pattern that history has seen before: open source wins the adoption war and then gets enclosed by whoever controls the silicon beneath it. If that is what happens here, the three companies will have given away MoveIt and Intrinsic Core in exchange for every new robot arm being built on a Dragonwing or a Jetson. Developers get a cheaper entry point either way.
The near-term beneficiaries are unglamorous and numerous: the machine shops, fab shops and mid-sized manufacturers that Intrinsic counts in the tens of thousands, and that have been told for years that automation is a project rather than a product. Open tooling does not fix the integration labour that makes robotics expensive. It removes the licensing conversation, which is the part that was stopping them asking.
Images
An Arduino Mega 2560 board, photographed at an angle rather than flat. The Arduino infinity logo and the MEGA 2560 model silkscreen are clearly visible, and the empty 28-pin socket shows no microcontroller installed. It is a classic hobbyist-class board, not one of the Qualcomm VENTUNO Q robotics boards named in the acquisition, but it represents the same developer base the announcement is courting. Wikimedia Commons, CC BY-SA.
Automation in a working laboratory: a small precision positioning stage reaches toward a rotary platform while a technician in an orange cleanroom garment stands behind. This is an illustrative example of the kind of instrument cell open tooling is aimed at — it is not an Intrinsic, Wandelbots, FANUC or Universal Robots installation. Wikimedia Commons.
The pipetting head of a laboratory liquid-handling robot, labelled P200/96, with a dense grid of transparent tips descending toward a microplate seated in the deck carrier. A real machine doing repetitive precision manipulation — the class of application that perception, planning and real-time control stacks are bought to serve. Wikimedia Commons.
References
- Qualcomm to Acquire PickNik to Advance the Future of Open Robotics and Physical AI — Qualcomm newsroom, September 23, 2026
- Introducing Intrinsic Core: An open source approach to Physical AI — Intrinsic, September 22, 2026
- ROSCon 2026: NVIDIA, Google, and Qualcomm Vie for Robot Software — Hannover Messe, October 2026
- NVIDIA Isaac ROS 5.0 brings AI agent capabilities to ROS developer ecosystem — Robotics 24/7, September 22, 2026