Seeing Machines unveils 3D human sensing for robots

Seeing Machines has published a technical paper on its real-time 3D human-understanding technology for robots and other Physical AI systems. The research focuses on the company's Human Mesh Recovery system.

The paper explains how the technology reconstructs a detailed 3D representation of a person from a single camera view, allowing a machine to interpret location, posture and movement in real time on embedded hardware. The system runs at up to 180 frames per second on NVIDIA Jetson Thor.

The development expands the company's work beyond transport safety, where it is known for AI-based human sensing, into robotics and industrial automation. It comes as developers of machines designed to operate near people seek better awareness of human movement without relying on remote processing.

The system is intended for places where robots and people share space, including factories, hospitals and warehouses. In those settings, the ability to judge how a person is moving and where they are positioned can affect how quickly and safely a machine responds.

High-detail 3D human reconstruction has often required significant computing resources, limiting its use on smaller embedded devices. The model is designed to address that constraint by processing data locally rather than sending it to the cloud, reducing reliance on network connections when machines need to react immediately.

The paper also states that the Human Mesh Recovery model works with both standard RGB cameras and RGB-D cameras without increasing its parameter count. That could broaden the range of hardware on which the system can be deployed, particularly in industrial environments where camera setups vary.

Robotics push

The technology forms part of Seeing Machines' Human-Centred Physical AI Platform, launched as a broader framework for machine perception in three-dimensional environments. The platform draws on the group's existing work in computer vision, human behaviour analysis and safety-focused AI.

The move reflects a wider commercial effort to apply the company's human-sensing research beyond automotive, fleet, aviation, rail and off-road markets. Seeing Machines built its reputation on systems that monitor human attention and behaviour in transport and is now positioning that experience for machines operating in industrial spaces.

For robotics companies, one challenge has been enabling machines not only to detect a person's presence but also to interpret how that person is moving as an interaction unfolds. The research focuses on that step, aiming to give robots a more detailed understanding of human motion from limited visual input.

Embedded deployment is a key part of the pitch. Running models directly on local hardware can be important in factories, healthcare settings and warehouses, where latency, connectivity and privacy concerns can shape how AI systems are adopted.

Company view

Paul McGlone outlined the rationale for the work in comments released with the paper.

"For robots to work safely and effectively alongside people, they need to understand more than simply whether someone is present. Our HMR technology delivers a compelling combination of accuracy, latency and model size on embedded hardware, bringing detailed 3D human understanding into real-world environments. This research demonstrates how Seeing Machines' 25 years of human-sensing expertise can support the next generation of Physical AI," said Paul McGlone, Chief Executive Officer, Seeing Machines.

Seeing Machines, founded in 2000 and headquartered in Australia, operates across Australia, the United States, Europe and Asia. Its products span automotive, commercial transport, aviation, robotics and industrial automation, with a focus on systems that help machines interpret people and the environments around them.

The new paper positions Human Mesh Recovery as a core part of that strategy, linking the company's established transport-focused work with its newer ambitions in robotics. Its central claim is that high-fidelity 3D human reconstruction can now be delivered in real time on embedded hardware previously seen as unsuitable for that level of scene understanding.