Seeing Machines is no longer merely exploring robotics. It has built a platform designed to become the human-perception layer inside the Physical AI ecosystem.

That may sound like a bold claim for a company best known for monitoring drivers, but Seeing Machines’ latest robotics demonstration suggests the strategy is now tangible.

The company has published a technical paper demonstrating its Human Mesh Recovery technology running on NVIDIA’s Jetson Thor robotics platform. HMR can reconstruct a detailed 3D representation of a person from a single camera, allowing a machine to understand human position, posture and movement in real time. On Jetson Thor, Seeing Machines says it can deliver up to 180 frames per second while combining high accuracy with a relatively compact model.

The significance is not simply that SEE appears to match the performance of much larger models such as Meta’s SAM 3D Body. It is that SEE is achieving this level of human understanding with a much smaller, faster model designed to run directly on the embedded hardware inside a robot.

A robot operating in a factory, warehouse, hospital or home cannot simply know that a human is nearby. It needs to understand where that person is, what they are doing and, ultimately, what they are likely to do next in real time. Seeing Machines’ new Physical AI platform is designed around precisely this problem, with capabilities spanning human detection, pose and movement estimation, behavioural understanding, prediction and support for safe interaction.

And this is no longer just a research project. Seeing Machines launched its Human-Centred Physical AI Platform in August, has demonstrated its technology on a Unitree G1 humanoid robot and is recruiting robotics engineers with experience in physical robots, perception, control, motion planning and human-robot interaction.

That points towards an important change in the investment story.

Seeing Machines does not need to become a robot manufacturer. Its opportunity may be considerably more valuable: becoming the human-perception layer that sits inside robots made by other companies.

The company’s platform is designed to work across multiple robot embodiments and hardware configurations. That opens the possibility of SEE supplying the intelligence that allows different types of robots to understand and interact with humans.

For a company that has spent more than two decades teaching machines to understand humans, this is a logical extension of its existing expertise rather than a completely new direction.

And it makes Seeing Machines increasingly interesting to another group of potential partners: chip companies competing for market share in Physical AI.

If you’re NVIDIA, Ambarella or another edge-AI processor company trying to get your silicon into the next generation of robots, an efficient human-perception model that makes your processor demonstrably useful for Physical AI is a very valuable piece of the ecosystem.

That makes SEE not merely a potential robotics supplier, but a human-perception platform that chip companies will want inside their Physical AI ecosystems.

The writer holds stock in Seeing Machines.