By Ivan Kesic

Iranian researchers recently secured a second consecutive victory in the Sussex-Huawei Locomotion Challenge, demonstrating how pretrained artificial intelligence models can be adapted to recognize human movement and transportation modes from smartphone sensors.

A research team from the University of Tehran won first place in the 2026 Sussex-Huawei Locomotion Challenge (SHL), an international competition focused on recognizing human movement and transportation modes using sensor data.

Led by Associate Professor Bagher Babaali of the university's Faculty of Mathematics, Statistics and Computer Science, the team became the challenge's champion for the second consecutive year.

A member of the team described the approach as an experiment with five pretrained time-series foundation models, using a global linear fusion strategy to combine their predictions in the final submission.

The achievement is noteworthy because it concerns a practical and increasingly important area of artificial intelligence: extracting meaningful information about human activity from the continuous streams of data generated by everyday electronic devices.

Rather than relying primarily on cameras or manually entered information, this approach uses motion sensors to infer whether a person is walking, running, cycling, traveling in a vehicle or remaining stationary.

The competition offers a useful illustration of how AI research is moving beyond systems that generate text and images toward models capable of interpreting measurements from the physical world.

It also highlights a central question in contemporary machine learning: can models trained in advance on broad collections of data be adapted to new tasks without the expense of retraining them from scratch?

What is the Sussex-Huawei Locomotion Challenge?

The Sussex-Huawei Locomotion Challenge is associated with research organized by the University of Sussex in the United Kingdom and supported by Huawei, a leading global provider of information and communications technology (ICT) infrastructure and smart devices

It forms part of the wider research community working on ubiquitous computing, wearable technology and mobile sensing. The 2026 edition was held in Shanghai, China, alongside activities associated with the UbiComp/ISWC research conference.

The challenge centers on human-activity recognition, a field of computer science concerned with identifying what a person is doing from data collected by electronic sensors.

Although the term can encompass activities such as sitting, climbing stairs, exercising or performing household tasks, the SHL challenge has a particular emphasis on locomotion and transportation.

Its eight categories are stationary activity, walking, running, cycling, traveling by car, traveling by bus, traveling by train and traveling by metro or subway.

Distinguishing between them is not always straightforward. A smartphone in a moving car may record vibrations and changes in acceleration that resemble those produced by other forms of transport.

Walking and running generate related but distinguishable motion patterns, while cycling can produce sensor signals that vary according to road conditions, speed and how the phone is carried.

The SHL dataset was developed to support research into these problems. The original dataset paper describes extensive real-world recordings collected using smartphones carried at different body positions, including the hand, torso, hips and a bag.

The phones recorded movement and other sensor information during daily activities and journeys in the United Kingdom.

The dataset is valuable because it captures the imperfect conditions encountered outside a laboratory. A phone can rotate, move inside a pocket, be held in the hand or be placed in a bag.

Road surfaces, vehicle suspension, walking speed and individual movement patterns all introduce variability. An algorithm that performs well only when a phone is held in one fixed position may be much less useful in everyday life.

The challenge therefore tests more than the ability to recognize a simple pattern under controlled conditions.

It explores whether machine-learning systems can identify locomotion and transportation modes despite changes in sensor readings and the circumstances in which those readings are collected.

Why smartphone motion sensors matter

Modern smartphones contain several sensors that can be used to analyze movement. An accelerometer measures acceleration along three axes, recording changes associated with movement, vibration and the effect of gravity.

A gyroscope measures angular velocity, helping describe how the device rotates. A magnetometer measures the surrounding magnetic field and can provide information relevant to orientation.

When these measurements are collected repeatedly, they form a time series: a sequence of observations ordered in time. The pattern across many observations can reveal information that an individual reading cannot.

Walking, for example, typically produces repeated movement patterns associated with successive steps. Running generally involves a different rhythm and intensity. A vehicle journey may generate a mixture of sustained motion, vibration, acceleration and braking.

Machine-learning models can learn to associate such patterns with known activity labels. Once trained and evaluated, a classifier can receive a new sequence of sensor readings and estimate which activity category best explains it.

The difficulty lies in generalization. A walking pattern recorded from one person may differ from that of another person. A phone carried in a trouser pocket may produce a different signal from the same phone held in the hand.

A bus moving along a smooth road may produce different measurements from one traveling over uneven pavement. Train and metro journeys can also share characteristics, making them difficult to distinguish in some circumstances.

These variations explain why human-activity recognition remains an active research area despite the widespread availability of smartphone sensors.

The challenge is not simply to identify movement, but to build systems that continue to work when the person, device placement or surrounding conditions change.

The role of time-series foundation models

The University of Tehran team's approach drew on time-series foundation models. The term refers to models pretrained to learn useful patterns or representations from sequential data.

The broader foundation-model concept became widely known through large language models, which learn statistical relationships in text before being adapted to tasks such as summarization, translation or question answering. Time-series models apply related ideas to measurements that change over time.

A time series might represent electricity consumption, weather measurements, industrial-machine vibration, financial observations or motion-sensor readings.

These applications differ substantially, but they share a common feature: the order and relationships between measurements contain information about the underlying process.

A model pretrained on time-series data may learn reusable representations of recurring patterns, trends or relationships.

Researchers can then adapt those representations to a more specific task, such as classifying a smartphone sensor sequence as walking, cycling or traveling by train.

This is potentially useful because assembling a large, carefully labeled dataset for every new application can be expensive. Labeling movement data may require participants to carry devices while their activities are recorded and verified.

Other applications may require specialist instruments, expert annotation or lengthy monitoring periods. A pretrained model offers a possible way to reuse knowledge acquired during earlier training rather than starting from zero for every new task.

However, time-series foundation models are not interchangeable with large language models. Their inputs, architectures, training objectives and intended uses can be different.

Nor does the word foundation guarantee that a model will perform well on every dataset. Its usefulness depends on factors such as the data used during pretraining, the compatibility of its input format with the target task and the method used to adapt it.

The 2026 SHL challenge provided a setting in which researchers could investigate precisely these questions.

According to a public statement by a member of the University of Tehran team, the researchers explored five frozen time-series foundation models, trained classifiers on top of their representations and combined their predictions using a global linear fusion strategy for the final submission.

Combining models rather than retraining them from scratch

The strategy is significant because it separates the general-purpose representation learned by a pretrained model from the smaller component responsible for the final classification.

In a typical frozen-model approach, the pretrained model's internal parameters remain unchanged. The model processes the incoming sensor sequence and produces a representation of its characteristics. A task-specific classifier is then trained to map that representation to the categories relevant to the competition.

This arrangement can reduce the amount of training required for the new task. It can also make it easier to compare different pretrained models, since researchers can evaluate their usefulness without undertaking the full computational expense of retraining each one.

The method does not eliminate training altogether. The classifier or other task-specific components still need to be fitted using labeled examples, and the complete system must be evaluated against data that were not used to train those components.

The quality of the result depends on both the pretrained representation and the way the downstream classifier is designed.

The second element of the approach was model fusion. Rather than relying exclusively on a single model, the team combined predictions from the five pretrained models it had explored.

An ensemble can sometimes improve performance when its component models make different kinds of errors.

If one model handles certain motion patterns well while another is better at distinguishing difficult categories, combining their outputs may produce a more reliable final prediction.

The advantage is not automatic. Models that make nearly identical errors may offer little benefit when combined, while a poorly designed ensemble can add complexity without improving the result.

Researchers must therefore evaluate the combined system against appropriate benchmarks rather than assume that adding more models will necessarily make it better.

The team's use of all five pretrained models in the final fusion is important. Rather than discarding models that performed less strongly on their own, the researchers combined their outputs through a linear fusion strategy, allowing the strengths of different models to complement one another.

This is a practical research approach: test a range of available tools, assess their suitability for the target problem and combine them in a way that produces a stronger final system than any single model could deliver alone.

Why the result has international relevance

The importance of the achievement extends beyond the competition itself. Human-activity recognition has potential applications in transport planning, mobile computing, health-related monitoring, fitness systems and context-aware digital services.

A smartphone that can distinguish walking from traveling by bus or train may help an application estimate how a person moves through a city.

Such information could support journey analysis, transport-mode classification, travel diaries and mobility research. At an aggregated and appropriately privacy-protected level, mobility data may also contribute to studies of travel behavior and urban transport systems.

Wearable devices could use related methods to identify patterns in physical activity, support exercise monitoring or help researchers analyze changes in movement. In industrial settings, similar approaches to time-series classification can be applied to machine vibration, equipment monitoring and other sensor-based tasks.

These are related applications of the underlying methods, not proof that the University of Tehran system has already been tested or validated for each of them.

The potential benefits are accompanied by important limitations. A transportation classifier is not automatically a medical diagnostic tool.

Inferring that someone is walking or running is different from detecting a disease, assessing fall risk or evaluating rehabilitation progress. Clinical applications require appropriate validation, representative participant groups and evidence that the system performs reliably in the intended setting.

The same caution applies to transportation systems. A model's performance on a research benchmark does not by itself establish that it can operate reliably across all smartphone brands, operating systems, countries, road conditions and patterns of human movement. Deployment may require additional testing, calibration and adaptation to new data.

Nevertheless, the competition provides a meaningful research environment because it asks teams to solve a defined technical problem using sensor data under common evaluation conditions.

A strong result indicates that the submitted approach performed competitively within that framework. It does not establish universal superiority across every possible application.

What the victory demonstrates

The University of Tehran team's 2026 result marks its second consecutive victory in the challenge. This indicates sustained performance across successive editions rather than a single isolated result.

That distinction matters in machine learning, where a promising outcome on one benchmark can sometimes be difficult to reproduce.

Repeated success provides stronger evidence that a team has developed useful expertise in data preparation, model selection, experimental design and evaluation.

The result reflects a genuine technical achievement: the team identified which pretrained time-series models were best suited to the task, designed an effective classification pipeline and combined predictions in a way that outperformed competing submissions. That is a meaningful contribution to a demanding international competition.

For Iran's research community, the consecutive victory is a notable achievement in an internationally organized technical competition.

It shows that Iranian researchers are working at the frontier of applied artificial intelligence and can compete successfully with teams from around the world.

The scientific value of the result ultimately lies in the method and the evidence supporting it. Research competitions are useful precisely because they allow different approaches to be compared on a shared task.

The University of Tehran team's success adds a valuable data point to that comparison and points toward the kind of approach that works well for sensor-based activity recognition.

A contribution to the expanding field of sensor-based AI

The University of Tehran team's victory illustrates a broader development in artificial intelligence: the growing use of pretrained models to interpret signals generated by the physical world.

While generative AI has attracted much public attention, systems that analyze sensor measurements may have equally practical significance in less visible areas of computing.

Recognizing transportation modes from smartphone data requires a combination of signal processing, machine learning, experimental design and an understanding of how human movement appears in sensor measurements.

The reported use of several pretrained time-series models, followed by a carefully selected fusion, shows how researchers can approach that problem without necessarily retraining every model from scratch.

For the wider field, the result's significance will grow as the team's approach is documented and others build on it. The method offers a template for adapting pretrained time-series models to new sensor-based tasks efficiently and effectively.

The next step is to examine what the team did, why its combination of models worked, how well it performed against competing submissions and whether the method can be extended beyond the benchmark.

Those are the questions that can turn a competition result into a lasting contribution to applied artificial intelligence—and the University of Tehran team has positioned itself well to lead that effort.