Algorized and ZHAW Study Sensor Fusion for Predicting Human Movement Near Robots
- Hannah Shapiro
- 498 reads
A two-year program will combine radar, Wi-Fi and optical signals to estimate people’s positions and paths for collaborative robot control.

Sensor fusion moves upstream of robot control
For a robot sharing space with people, a useful detection system has to do more than report a position. It must estimate how reliable that position is and whether the person is moving toward the robot. A two-year research program from Algorized and the Institute of Signal Processing and Wireless Communications (ISC) at the ZHAW School of Engineering will investigate how to produce those estimates 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 and is funded by InnoSuisse. It builds on earlier work by Algorized and ISC involving ultra-wideband (UWB) radar. The research is intended to connect sensor measurements to a robot controller through a fusion layer that evaluates inputs together.
That placement matters. In a conventional sensor chain, each device may first interpret its own readings and pass a result onward. The researchers will examine whether combining raw inputs before individual sensors make their own decisions can give the system a more complete view. The fusion layer is intended to weight each input according to confidence, then produce a calibrated estimate of a person’s location and likely path.
Different sensors cover different blind spots
The sensor types under study contribute different information. UWB radar and Wi-Fi sensing can help detect a person when a shelf or machine blocks a direct view. Millimeter-wave radar can maintain detection through dust, steam, glare and darkness. RGB and time-of-flight cameras can contribute spatial and semantic detail. The research will test how these inputs work together, including how their differing strengths affect the system’s estimate.
These complementary measurements could help a robot respond when one sensor has limited visibility or produces an uncertain reading. But combining sensors also makes confidence assessment a central technical problem: the controller needs an estimate it can interpret, not simply a larger collection of measurements. The program’s aim is to send the controller a graded instruction—continue, slow down or stop—based on the fused perception results.
Algorized currently processes raw data with a foundation model to detect human presence and classify whether a detected object is human or non-human. The research will test whether sensor fusion at the raw-input stage can support a broader, more holistic layer for machines that need to account for people nearby. The description of the program does not specify how the model will resolve conflicting sensor inputs, or what confidence thresholds will trigger each controller instruction; these are among the practical details that the research must address.
Four targets, with an edge-computing constraint
The work has four stated goals: detect people approaching a robot, distinguish living beings from non-living objects, track and predict each person’s path, and translate the result into a graded controller instruction. Algorized’s planned safety functions cover presence detection, localization and tracking. The inference target is under 100 milliseconds on standard industrial edge hardware.
That timing target puts the focus on processing close to the robot and its sensors. An estimate that arrives too late may no longer describe where a moving person is or where they are headed. The stated target gives the project a concrete performance objective, though the program description does not provide test results, hardware models or measurement conditions. Those details will be needed to judge how well the system performs in a specific installation.
The research also faces a data challenge. A model intended to work in facilities needs training data from real sites, where people move through changing work areas and encounter equipment, shelves and other obstacles. The researchers say that people who understand the underlying physics must interpret this data. That points to a need to connect model training with the behavior of the sensors themselves, rather than treating the readings as interchangeable inputs.
A further goal is transfer: the team wants a core model that can work across sensor types, vendors and sites without retraining for every deployment. If achieved, that could reduce the effort needed to adapt perception systems to different industrial settings. The program description sets this out as an objective, not as a demonstrated result; whether the model can generalize across those variations remains to be established.
Research in industrial settings
Algorized CEO and founder Natalya Lopareva said the collaboration will explore sensors and compute layers to help robots understand people in shared spaces. Marc Kuhn, ISC deputy head, said the team will bring the institute’s sensing research into real industrial sites and study both movement estimation and ways to measure its reliability.
For engineering teams, the key outcome to watch is how the researchers turn sensor agreement, uncertainty and predicted movement into a controller response within the under-100-millisecond target. The project is scheduled to run for two years from September 2026. Its practical value will depend on whether the fused estimates remain useful across real facilities, sensor configurations and changing visibility conditions.