SUMediPose is a high-precision, multimodal 2D–3D human pose estimation dataset developed to address the limitations of existing markerless motion capture systems. By combining gold-standard marker-based Vicon motion capture with synchronized multi-camera video recordings, the dataset provides anatomically accurate ground truth data across a range of strength and conditioning exercises performed at varying speeds. With over 3,400 recordings and millions of annotated data points, SUMediPose enables researchers to train and validate more reliable pose estimation models. Its inclusion of detailed calibration parameters and aligned 2D–3D data offers flexibility for advanced research applications. The dataset ultimately supports improved biomechanical analysis and the development of more accurate, accessible motion capture solutions in medical and sports science fields.

This research introduces a low-cost, wireless dendrometer system that enables large-scale tree growth monitoring without the high cost of traditional equipment. By combining modular sensors, solar power, and long-range LoRaWAN communication, the system allows continuous, remote data collection across vast plantations. A key innovation is the use of a digital twin, which corrects temperature-related measurement errors through modelling rather than expensive hardware. Field and lab tests show that the system achieves high accuracy and reliability, comparable to commercial solutions costing significantly more. The result is a scalable, open-source platform that makes precision forestry more accessible, allowing researchers to monitor many more trees, improve experimental design, and gain deeper insights into environmental and biological processes.

This blog explores the development of an autonomous sailboat designed to navigate and operate independently over long distances using wind power. The research presents a practical, open-source approach by converting a commercially available sailboat into a smart vessel through a combination of physics-based simulation and advanced control systems. By creating a validated digital twin and integrating motion and guidance controls, the project demonstrates reliable autonomous navigation, including complex maneuvers like tacking. The work lays a strong foundation for future improvements in hardware and software, advancing the potential of low-cost, long-range ocean monitoring and exploration.

Professor Arnold Rix’s inaugural lecture highlights South Africa’s unique opportunity to lead in renewable energy, while addressing the critical challenge of grid stability in a low-carbon future. His research focuses on bridging the gap between clean energy generation and reliable power delivery through advanced modelling, hybrid systems, and real-world industry collaboration. By combining solar, wind, storage, and intelligent forecasting, his work aims to create resilient, decentralised energy systems that support both national infrastructure and broader economic development.

This research focuses on addressing the urgent need for faster, more accessible tuberculosis (TB) screening in South Africa. Dr Stephan Schoeman developed a portable, low-cost electrochemical biosensor designed to detect TB-related biomarkers quickly at the point of care, reducing reliance on complex laboratory infrastructure. Through iterative prototyping, the team engineered a compact, disposable device using carbon nanofibre electrodes and a layered design with built-in test validation. The final sensor (Sensor M) demonstrated high sensitivity across clinically relevant ranges by detecting C-reactive protein (CRP), delivering reliable results within minutes. The biosensor’s affordability, ease of use, and adaptability make it a promising tool for use in resource-limited settings such as mobile clinics. Beyond TB, the platform has potential applications in diagnosing other diseases, supporting broader public health efforts. Next steps include clinical validation, scaling up manufacturing, and expanding the device to detect multiple biomarkers simultaneously, further improving diagnostic accuracy and impact.

The blog explores how South Africa’s transmission network can be better utilised to support the rapid growth of renewable energy. While the Integrated Resource Plan (IRP) sets ambitious targets for wind and solar expansion, grid capacity—especially in regions like the Northern Cape—remains a key constraint. Traditional deterministic methods for calculating network capacity rely on worst-case scenarios, which often lead to overly conservative limits. The blog highlights that probabilistic methods offer a more realistic approach by considering a range of possible conditions. This allows for improved use of existing infrastructure without unnecessary and costly upgrades. The research ultimately shows that probabilistic techniques can provide better insight into true network capacity and help bridge the gap between South Africa’s renewable energy goals and the practical limitations of the grid.

This blog explores a vision-based localisation system designed to improve the autonomy of inspection drones (UAVs), especially in environments where GPS is unreliable or unavailable. The research focuses on markerless localisation, where drones use onboard cameras to detect natural features in their environment instead of relying on artificial markers. A 3D feature map is created from recorded data, and these features are clustered into “natural markers” that the drone can recognise in real time to determine its position and orientation. To improve accuracy and stability, the system combines vision-based data with inertial measurement unit (IMU) inputs, allowing continuous and reliable tracking. Overall, the approach enables more flexible, efficient, and autonomous drone inspections across industries, reducing the need for manual setup while improving performance in complex environments.

This blog explores how cooperative multi-agent reinforcement learning (MARL) can be improved in complex 3D environments where agents receive limited information and sparse rewards. The research adapts a single-agent algorithm (QMIX) for teamwork scenarios using a Portal 2-inspired simulation, where agents must coordinate actions to solve puzzles. To improve performance, several enhancements were introduced, including parallel experience collection, prioritised learning from important events, and memory-based learning using LSTM. A curriculum learning approach—starting with simple tasks and gradually increasing complexity—proved essential for success. The results show that agents can learn effective cooperation, even in challenging environments, and can transfer knowledge to new tasks. Overall, the study demonstrates that with the right training strategies, MARL systems can enable better coordination in real-world applications like robotics, autonomous vehicles, and simulation environments.

This blog highlights a new international research collaboration aimed at improving the stability of South Africa’s power grid. Stellenbosch University, alongside UK and local partners, is contributing to a UK-funded study that uses data-driven methods, statistical physics, and machine learning to better understand and manage electricity supply and demand. The project focuses on addressing ongoing challenges such as load shedding, system instability, and reliance on diesel generators. By developing predictive models and exploring alternatives like solar and battery solutions, the research aims to reduce outages and support a more sustainable energy future. In addition to technical outcomes, the initiative also promotes skills development, collaboration, and knowledge exchange between South Africa and the UK, helping strengthen long-term capacity in energy research.

This blog explores a new approach to improving place recognition in vision-based SLAM systems by incorporating semantic information. Traditional methods rely mainly on geometric features like points and edges, but often ignore what those features represent in the real world. The research introduces SeM2DP, a method that combines 3D spatial data from stereo cameras with object recognition from a neural network. This creates a compact “fingerprint” of a scene that includes both structure and meaning. By doing this, robots can recognise previously visited locations more accurately, improving loop closure and reducing mapping errors. Testing shows that SeM2DP outperforms traditional methods in accuracy while using less storage, although it currently runs slower due to the added semantic processing. However, in systems already using AI for object detection, this added step comes at little extra cost. Overall, the approach demonstrates that adding context and meaning to visual data significantly improves how robots understand and navigate their environment, leading to more reliable and accurate mapping.