Algorized and ZHAW Study Sensor Fusion for Human-Aware Collaborative Robots

Collaborative robots need more than a signal that a person is nearby. Their controllers need an estimate of where that person is going, how reliable the estimate is, and what response is appropriate. A two-year research program from Algorized and the Institute of Signal Processing and Wireless Communications (ISC) at ZHAW Zurich University of Applied Sciences will investigate how to produce that information by combining wireless and optical sensor data.

The program, titled “Real-Time Human Perception via Edge-AI Wireless Sensor Fusion for Collaborative Robotics”, began in September 2026. InnoSuisse funds the work, which builds on earlier research by Algorized and ISC involving ultra-wideband (UWB) radar.

Combining signals before sensors make separate decisions

The researchers plan to combine raw wireless and optical signals in a fusion layer between a robot’s sensors and its controller. The system is intended to assign confidence to each input, then produce a calibrated estimate of a person’s position and predicted path. The target is a useful view of the surrounding space that accounts for the strengths and limits of different sensing methods.

That approach differs from simply collecting the final output of each sensor. Algorized currently processes raw data with one foundation model to detect presence and classify whether a detected object is a person or something else. The research will test whether combining raw inputs before each sensor makes its own decision can give a machine a more unified interpretation of its surroundings.

The distinction matters because sensing conditions vary across a facility. A camera can provide spatial and semantic detail, but its view may be affected by darkness or glare. UWB radar and Wi-Fi sensing may help detect a person who is out of direct view behind a shelf or machine. Millimeter-wave radar can maintain detection through dust, steam, glare and darkness, as described in the program. Fusion could let the system use complementary evidence when one input is less informative.

Combining sensors also creates a calibration problem: inputs may differ in coverage, timing and reliability. The program will examine how to weigh those inputs and measure confidence in the resulting estimate. Its stated goal is a model that can transfer across sensor types, vendors and sites without retraining for every deployment. Whether that transfer works in practice is one of the questions the research is set up to examine.

From detection to a graded controller response

The work has four aims: detect people approaching a robot, distinguish living from non-living objects, track each person and predict their path, and provide the robot controller with a graded instruction. The specified response levels are continue, slow down or stop. That structure ties perception to a practical control decision: the system must estimate not only whether a person is present, but whether the robot’s next action should change.

The program sets an inference target of under 100 milliseconds on standard industrial edge hardware. That target puts the focus on processing near the robot, where sensing and control decisions can be made without relying on a distant computing system. The source does not give details about the hardware configuration, sensor sampling rates or test conditions, so the target should be read as a research objective rather than a reported performance result.

Algorized’s planned safety functions cover presence detection, localization and tracking. Together, these functions are intended to supply the controller with information about a person’s location and movement. The research will also investigate how the reliability of that information can be measured, according to Marc Kuhn, ISC deputy head at ZHAW’s School of Engineering.

Training data must reflect real facilities

A model intended to work across different sites needs training data that captures the variety of movement and sensing conditions found in real facilities. The researchers say the data must come from settings where people move through active environments, and that specialists who understand the underlying physics must interpret it. This makes data collection and interpretation part of the engineering challenge, alongside model design and real-time processing.

Site-to-site transfer is a practical test of the proposed approach. A system that needs retraining for every sensor arrangement or facility would require additional work at each deployment. The project’s aim is to reduce that dependence by developing a core model that can carry across sensor types, vendors and locations. The program’s results will show how far that goal can be met under real industrial conditions.

Algorized CEO and founder Natalya Lopareva said the collaboration will explore sensors and compute methods to help robots understand people in shared spaces. Kuhn said the ISC research will combine radar, Wi-Fi and optical signals to identify approaching people, estimate position and movement, and investigate prediction of where a person may go next.

The central engineering question is whether a fused estimate can be both timely and dependable enough to guide a robot’s response. The two-year study will test the sensing and inference approach in industrial settings, with the under-100-millisecond target and cross-site transfer as concrete measures to watch.