Let’s start with what it is. It’s easy to “Explain Like I’m Five”: Physical AI means Robots. Broadly speaking, these robots can take all sorts of shapes and sizes. They might be humanoid servants helping in your home, autonomous taxis driving you around, delivery drones flying to your neighbourhood, or huge factory machines that build your furniture.
Physical AI implies the integration of “smarts”, usually with AI models or computer vision, in machines that can sense, decide and act in the real world. Physical AI systems share the trait that they must operate properly, whether or not they’re connected to the cloud. These robots need the most crucial AI services “at the edge” (running on the embedded target hardware), and cannot, or should not, depend on cloud servers for proper operation. Furthermore, most Physical AI applications are mission-critical such that if such a system crashes or malfunctions, it might harm people or destroy property, or, generally speaking, cause significant harm.
Okay, clear enough. So why are you hearing more about QNX and Physical AI all of a sudden? Actually, this domain isn’t at all new for us. Automakers, factory automation OEMs, general robotics, and medical surgery equipment makers have been depending on us for decades. And, to one level or another, they’ve been building smart products, using the progenitors of today’s AI technology, on QNX-based systems.
But what is relatively new for QNX is how seriously we’re taking the AI revolution. To be clear, I’m not saying we’re adding AI-generated code to our process or products - something that should assure our longstanding customers who value reliability, safety, and stability. Rather, I’m talking about ensuring that those building AI-assisted systems—products using LLMs, neural nets, AI models, and ML algorithms—will have access to industry-standard and cutting-edge software enabling those “smart” features on QNX.
How do we provide iron-clad commercial guarantees and the process rigor for AI technology, just like we do with our commercial products? Well, quite honestly, we can’t. AI is moving far too fast for us to attempt to safety certify or ISO-process those bits into submission. The landscape is ever-evolving: there are many players, hundreds of libraries and frameworks, and millions of commits each week.
But while we can’t necessarily tame the AI tiger into functionally safe territory (yet), we’ve architected an approach that lets us offer real value to the industry and developers in general in the meantime. That is, we have our QNX Everywhere program act as the tip of the spear and feed into our Physical AI strategy. QNX Everywhere is our program that provides free access to QNX software for non-commercial use. (In other words, we only want your money when you’re making money.)
The main products within QNX Everywhere – QNX Software Development Platform 8.0 and QNX Hypervisor 8.0, our operating system and tools – are identical to our commercial product. But there are tons of open-source projects that we (and others) have ported to QNX and are readily available for you to use. That gives us short-term agility for demos, prototyping, and proof-of-concept. As the market starts to pick AI software tech winners and losers, we can then make sure this software is available and supported on QNX through us or trusted 3rd party partners. In Silicon Valley parlance, QNX Everywhere gives us a tiny taste of “move fast and break things” – at least, as much as a reliability-based brand like ours can tolerate!
Let me make this concrete. My colleagues recently went to Boston for the Robotics Summit & Expo. There, in addition to several excellent customer demos that use QNX under the hood, we were showing off two open-source demos we created. The first was our “infamous” baby robot arm with added camera, pose recognition, and inverse kinematics. In English, this is a project that anyone can DIY on a Raspberry Pi to recognize their gestures and translate them onto a little plastic robot arm.
The second demo was in a similar vein for digital twin tech. It ran on an Intel NUC with an Nvidia RTX; it recognized your body position, converted what it saw into a “skeleton”, pushed that geometry through ROS, and visualized the result with a 3D model robot with RVIS. That part was running on an Ubuntu system in the QNX Hypervisor on the same NUC. In short: a little on-screen robot copies your movements.
These demos uncovered another thing that makes our platform compelling for Physical AI. QNX is real-time, and while that might not matter for lots of applications for real-time determinism and performance robotics it absolutely does.
A Raspberry Pi is a great little chunk of silicon, but it’s not a computational powerhouse. So, we used a Hailo HAT (Hardware Attached on Top) to offload pose detection, which is bounded by PCI performance. Since we respond immediately and consistently to the huge amount of data constantly flowing through that pipeline, we see a significant performance boost over using Linux in the exact same use case. Same thing with the Intel NUC digital twin demo. The frame computation time is around 15 milliseconds on QNX; for similar bottleneck reasons, it’s nearly an order of magnitude higher on Linux. (If we’re going to get into the weeds, even against a PREMPT_RT Linux distro, we’ll still outperform it.)
We’ll be making these available as open-source demos for people to build, expand, and learn from.
Beyond the demos and prototypes, we love being the platform that’s enabling commercial products with Physical AI – with real-time, determinism, reliability, and safety/security certified software. And with QNX Everywhere, we’re ensuring that we stay at the cutting edge of that work. For the technologists who either want to know exactly what we support, what we’re adding, or how to contribute, stay tuned! I’ll be putting out another blog shortly to address that.