Seeing Machines has published a technical paper outlining a method for real-time 3D human understanding on embedded hardware, a capability aimed at improving safety and interaction in systems described as “Physical AI”, including robots operating around people.
The paper describes the company’s Human Mesh Recovery (HMR) technology, which reconstructs a detailed 3D representation of a person from a single camera view so a machine can interpret a person’s location, posture and movement in real time. Seeing Machines positions the technology for settings where people and robots work in close proximity, including factories, hospitals and warehouses.
According to the company, Physical AI systems need more than person detection to respond safely as interactions unfold, but high-fidelity 3D human reconstruction has generally required substantial computing resources, making real-time deployment on embedded devices difficult.
Seeing Machines said HMR is designed to address this by processing information locally at the edge, allowing systems to react in real time without relying on a cloud connection. The company reported performance of up to 180 frames per second on NVIDIA Jetson Thor hardware, arguing the results show detailed 3D reconstruction can be achieved on embedded platforms previously considered unsuitable for that level of scene understanding.
The paper also states the HMR model supports both RGB and RGB-D cameras without increasing its parameter count.
CEO Paul McGlone said the aim is to help robots interpret human motion in addition to presence, and that the work builds on the company’s experience in human-sensing.
Seeing Machines said HMR is a core capability within its Human-Centred Physical AI Platform, which it said was launched in August 2026, and that it plans to apply experience from automotive, commercial fleet, aviation, rail and off-road deployments to robotics and industrial automation.
You can read the full report here.
