IN Brief:
- Algorized and ZHAW have launched a two-year, Innosuisse-funded research programme combining UWB radar, Wi-Fi sensing, millimetre-wave radar and camera data.
- The proposed sensor-fusion architecture will estimate human position and movement, assign confidence to its measurements and inform robot-control decisions.
- Researchers target inference below 100 milliseconds on industrial edge hardware; neither that performance nor functional safety certification has been demonstrated.
Algorized and the ZHAW School of Engineering have launched a two-year research programme investigating how wireless and optical sensing can be combined to predict human movement around collaborative robots. Funded by the Swiss innovation agency Innosuisse, the project will develop a sensor-fusion architecture that estimates a person’s position and likely path while reporting the confidence associated with those measurements.
The programme, announced on 20 October 2026, began in September and builds on earlier work between Algorized and ZHAW’s Institute of Signal Processing and Wireless Communications. Its research will combine ultra-wideband radar, Wi-Fi sensing, millimetre-wave radar, RGB cameras and time-of-flight imaging, using edge AI to interpret measurements obtained under different environmental conditions.
The proposed system will sit between the robot’s sensing hardware and its motion controller. Rather than requiring each instrument to produce a separate classification before its information is combined, the researchers intend to investigate fusion of the underlying sensor data. The resulting processing layer would estimate human presence, distinguish people from other objects and predict movement around the machine.
Collaborative robotics presents difficult perception conditions because workers and equipment may occupy the same operating area. A person can move behind shelving, approach a robot from outside a camera’s field of view or become partially obscured by materials being handled. The sensor system must also distinguish these events from reflections, moving equipment and changes in the surrounding environment.
Optical cameras provide spatial detail when the relevant objects are visible. RGB images can support recognition of people and equipment, while time-of-flight cameras measure distance using the behaviour of emitted and returned optical signals. Their performance nevertheless depends on the particular sensor design, available illumination, surface properties and the extent to which the observed scene is obstructed.
Radio-frequency sensors provide measurements with different limitations. Ultra-wideband radar transmits signals across a relatively broad frequency range and analyses their reflections to obtain information about nearby objects. Under suitable conditions, this can provide evidence of movement where direct optical observation is restricted, although radio propagation depends on the surrounding materials and geometry.
Wi-Fi sensing uses changes in radio signals to infer aspects of human presence or motion. A person’s movement can alter the propagation paths between transmitting and receiving equipment, creating measurable variations in the received signal. Interpreting those changes requires calibration and signal processing because the same environment may contain numerous stationary and moving reflectors.
Millimetre-wave radar contributes further information about position and movement. Such systems may operate under lighting conditions unsuitable for ordinary cameras, although their performance remains affected by antenna configuration, target characteristics, signal processing and interference. None of the sensor types can be assumed to provide reliable coverage of every possible operating condition.
Combining these measurements introduces a separate engineering problem. The instruments may operate at different sampling rates, use different coordinate systems and provide estimates with different uncertainties. The processing system must associate observations from the same object, account for their timing and determine which measurements can be relied upon when the available information conflicts.
Algorized and ZHAW intend to assign confidence weights to sensor inputs and produce a calibrated estimate of human position and predicted movement. A confidence measure can identify when the available evidence is uncertain, allowing the receiving controller to distinguish a comparatively well-supported position estimate from one produced using incomplete or inconsistent observations.
The researchers will also investigate the use of a foundation model to process information from multiple sensing modalities. A common model may be able to identify patterns that are difficult to reproduce using independent classifiers, although this depends on the quality of the training data, the relationships between the inputs and the model’s behaviour in operating conditions not represented during development.
The programme defines four principal objectives: detecting people approaching the robot, distinguishing living from non-living objects, tracking and predicting individual movement, and supplying graded information to the controller. The proposed control responses are to continue operating, slow down or stop, depending on the predicted situation and the confidence of the measurement.
Those responses require integration with the robot’s established control and protective functions. A perception model can supply useful information about the surrounding environment without itself constituting a certified safety function. Any system relied upon to prevent hazardous contact must be assessed against the applicable machinery safety requirements and the complete behaviour of its sensing, processing and actuation chain.
Algorized and ZHAW are targeting inference latency below 100 milliseconds on standard industrial edge computing hardware. The figure is a research objective rather than a verified performance result. It also concerns inference rather than the entire time required for a robot to react, which includes sensor acquisition, communications, controller processing, drive behaviour and mechanical stopping.
Latency becomes particularly important when people and machinery are moving towards one another. During a delay between measurement and physical response, the separation distance may continue to decrease. The complete protective design must therefore consider movement speed, braking capability, measurement uncertainty and the consequences of a sensor or processor fault.
The project will use data from industrial environments to investigate the behaviour of the sensing system beyond controlled laboratory conditions. Machinery, reflective materials, moving objects and variable lighting can change the characteristics of the measurements, while the position of sensors relative to the working area influences what they can observe.
Training and validation across different installations will be necessary to establish whether the proposed fusion architecture can operate reliably when its environment changes. A model that performs well with one sensor arrangement may require further calibration or adaptation when deployed with different hardware, altered sensor positions or unfamiliar objects.
The research is also intended to examine whether a shared perception model can support several industrial sensing configurations without extensive retraining. No completed cross-site validation dataset, independently verified inference result or functional safety certification was disclosed with the launch announcement.
Over the two-year programme, Algorized and ZHAW will investigate the accuracy, latency and uncertainty of the combined sensing approach. The resulting evidence will determine whether its movement predictions and confidence estimates are sufficiently reliable to inform subsequent collaborative robot control development.
