Phystech.Genesis has opened applications for the third archaeology contest as part of the Up Great "Expedition. Data Science" technology contest (NTF Foundation). The stage called KOZ No. 3 "Integration" requires teams of 2–5 people with machine learning and computer vision skills to find objects not only in the terrain and on the surface, but also underground, by combining heterogeneous sensor data. Applications are accepted until September 30, 2026, the stage's prize fund is 30 million rubles, and the dataset is being prepared at the Gnezdovo archaeological complex.

What happened

Data Science", which is run by the NTF Foundation. Individuals and legal entities can participate: applications are submitted by teams of 2–5 people with machine learning and computer vision skills. Work on solutions is scheduled for October–November 2026 and will take place on the Phystech.Genesis online platform. The stage's prize fund is 30 million rubles. The main difference from previous stages: models must find objects simultaneously in the terrain, on the surface, and underground, in the cultural layer and bedrock. The dataset is being prepared by a team led by archaeologist Vasily Novikov at the Gnezdovo archaeological complex, where finds about 1,000 years old are documented using scanning, geophysics, thermography, and robot dogs; participants will be provided with both processed and raw data.

Context

"Expedition. Data Science" is a series of technology contests by the NTF Foundation, in which the "Detection" and "Scanning" stages have already been held. The essence of the new task is integration: fusion of heterogeneous sensor data, including aerial photographs, satellite images, lidar, magnetometry, and ground-penetrating radar, to detect objects at depths of up to 5 meters. This is no longer classic detection on images, but a multi-modal fusion task with sparse and noisy geophysical labeling, where the detector architecture is not what matters, but the alignment of modalities: resolution, georeferencing, calibration, and the fundamentally different physics of magnetometry and ground-penetrating radar signals. The Gnezdovo data is valuable also because there are almost no labeled samples for subsurface probing available in the public domain, and access to raw measurements allows the entire preprocessing to be reproduced, not just tuning a model on ready-made datasets.

Why this matters for the industry

For the industry, this is one of the largest public ML contests in Russia, and its task essentially sets a product template of "heterogeneous sensors → preprocessing → map of anomalies in the terrain, on the surface, and underground." Such integration pipelines are directly applicable in archaeological surveying, geotechnics, and remote sensing, that is, everywhere you need to look under the surface without excavations. For ML teams, the main value may not be in the prizes, but in access to a rare labeled sample for sensor fusion from the Gnezdovo complex; those who only have experience in detection on photographs will find a reason to build out a sensor alignment pipeline. If the organizers publish the evaluation protocol and results, by the end of November 2026 it will be clear whether modality integration provides a real gain over single-channel approaches — there is no public data of this kind yet, and without an open protocol the results will remain internal to the contest.

Why this matters for users

A team of 2–5 people with computer vision and machine learning skills just needs to register on the expds-platform.upgreat.one platform: applications are accepted until September 30, 2026, inclusive, until 23:59 Moscow time. Participation in the previous "Detection" and "Scanning" stages is not required — you can enter KOZ No. 3 directly and compete for a share of the prize fund. Participants will have the opportunity to work with real field data collected at an actual archaeological complex: such experience is a rarity in itself and strong material for a portfolio. Practical step: it makes sense to prepare a raw geophysical data preprocessing pipeline in advance, before the stage starts in October, and if the team only has CV specialists, it is worth inviting a partner with expertise in geophysics or archaeology.

What is still unknown / limitations

Open materials currently lack a quality metric, train/test protocol, labeling description, and baseline solutions, so it is impossible to judge the reproducibility and comparability of results — this is a key open question for the organizers. The claim about the transferability of the same pipeline to geotechnics, utility search, or construction site monitoring remains a plausible hypothesis, not a proven fact: the physics of magnetometry and ground-penetrating radar signals and the background structure in other domains are different. Long-term scenarios — turning the Gnezdovo dataset into an industry benchmark for multi-modal subsurface detection and pilots in adjacent fields — depend on an open evaluation protocol and the continuation of the contest series. Currently, this is a contest stage, not a product release: the organizers have not published any API, prices, or latency data.

Sources

Author

Look at AI, editorial team