IEEE Information Theory Society · Student & Outreach Committee
ESIT-D2I Competition – Charting the Path to Bifröst

The First ESIT-D2I Competition

Charting the Path
to Bifröst

Channel Charting · Robust Learning · Unlearning Noisy Contributions

A student competition at the intersection of data science, machine learning, and information theory — held in conjunction with ESIT 2026, Nordfjordeid, Norway.

See the Final Results

Competition concluded  ·  Prize pool: 1,000 EUR  ·  Winners announced: July 8, 2026

★ Final Results · July 8, 2026

They Charted the Path to Bifröst

Four teams crossed the bridge. Here is the final leaderboard.

🏆 Champions
KUDFI
1st Place
  1. 🥈 2nd Yggdrasil Embeddings
  2. 🥉 3rd Asgard Autoencoders
  3. · 4th Solo Skald

01 · ABOUTThe Challenge

This year's challenge is inspired by the Viking legend of Bifröst — the shimmering bridge connecting Midgard to Asgard, where Valkyries ride across the sky to bring fallen heroes to Valhalla.

Back then, reaching Asgard required divine intervention, a winged horse, and — ideally — having died heroically in battle. Today's Nordic explorers, however, have 5G coverage reaching roughly 99.7% of households in Norway.

So we ask a simple question: If somewhere between Midgard and Asgard there were at least a stable cellular connection, and you were given noisy, partial, and imperfect measurements of the wireless channel — can you chart your way to Bifröst?

Welcome to the first edition of the European School of Information Theory Data-to-Information (D2I) Competition, organized by the IEEE Information Theory Society Student and Outreach Committee. Participants will work with real-world Channel State Information (CSI) measurements, designing methods that are robust to label noise, feature corruption, and faulty data contributions.

02 · PLATFORMSWhere the Competition Lives

Throughout the competition, everything lived one click away: the Slack community for announcements and team-building, and the two Kaggle competitions hosting the data, baseline code, and live leaderboards. These platforms are now closed — they remain listed below as a record of the competition.

Community
Slack Workspace
The single hub for announcements, technical Q&A, team formation, and all competition discussion, centred on the #d2i-esit2026 channel.
Workspace closed
Task 1 · Main Challenge
Learning under Noisy Conditions
Data, baseline code, submission interface, and live leaderboard for the robust-localization task.
Competition closed
Task 2 · Advanced Challenge
Unlearning & Data Reliability
Data, baseline code, submission interface, and live leaderboard for the unlearning task.
Competition closed

03 · FORMATCompetition Structure

Phase 1 · In-Person

Tuesday, June 2 — ESIT 2026

  • Introduction to objectives and problem setting
  • Presentation of dataset, notebooks, and baselines
  • Open Q&A and team formation session
  • Coffee and snacks provided

↓  Slides from this session (PDF)

Phase 2 · Online

June 4 – June 25

  • Participants work remotely on the challenge
  • Development and submission of solutions via Kaggle
  • Ongoing support through the competition Slack

Final Session · Online

June 26 · July 7–8

  • Selected teams announced for presentations (June 26)
  • 10-minute presentations per team (July 7)
  • Official winners announced (July 8)

Open to all · Live online

Live Presentation Session

Selected teams present their approaches and results on channel charting, robust learning, and unlearning noisy data contributions — working toward better localization quality from real-world CSI measurements. Everyone is welcome to attend.

Date

July 7, 2026

Time

16:00 – 17:00 CEST (GMT+2)

Live session concluded

04 · TIMELINEKey Deadlines

June 3, 2026
Online Enrollment
June 4, 2026
Team Formation Finalised
June 25, 2026
Solution Submission Deadline
June 26, 2026
Selected Teams Announced for Presentations
July 7, 2026
Final Presentations
10 min per team
July 8, 2026
Winners Announced

05 · REGISTRATIONHow to Join

Registration is now closed. During the competition, participants registered through an online form, indicating whether they were joining as an individual or as part of a team, and whether they planned to attend ESIT in person.

Team formation was finalized on June 4, after the in-person ESIT session — giving participants an opportunity to meet and form teams on-site. Each team was required to include at least one member attending ESIT in person. Remote participants were assisted by the organizers in forming teams.

Registration closed

06 · PRIZESPrize Pool

🏆 Best Team
500 €
Overall winner
🥈 Second Place
250 €
🥉 Third Place
150 €
🌱 Junior Prize
100 €
Youngest team that impresses the judges

07 · DATASETData

Participants will use a dataset collected in a realistic wireless environment (the DICHASUS dataset, Arena2036 research campus) where multiple remote antenna arrays capture Channel State Information (CSI) measurements. CSI describes how a wireless signal propagates through the environment — capturing reflections, scattering, and attenuation — and can be seen as a rich "fingerprint" of location.

The dataset is divided into two components:

Full dataset and tutorials: dichasus.inue.uni-stuttgart.de. All data, utilities, and the submission interface are hosted on Kaggle — see the Platforms section above for the Task 1 and Task 2 competition pages.

🧪

We gratefully acknowledge Prof. Stephan ten Brink, Florian Euchner, and Phillip Stephan for building and documenting this dataset and making it freely available to the research community.

08 · TASKSCompetition Description

Task 0 · Foundations (discussed at ESIT)

Warm-Up: CSI Localization Basics

This preliminary task introduces the core elements of CSI-based localization and serves as a warm-up for the competition. Participants will explore basic localization (mapping CSI to positions via supervised learning), basic channel charting (low-dimensional representations preserving spatial structure), and trajectory-aware processing (temporal consistency across consecutive measurements).

Task 1 · Main Challenge

Learning under Noisy Conditions

In realistic scenarios, both labels and measurements are imperfect. Participants design models that remain accurate and robust when data quality degrades. Two types of corruption are considered:

  • Label Noise (Position Corruption): Position estimates drift over time due to IMU accumulation errors.
  • Feature Noise (CSI Corruption): CSI measurements altered through phase offsets or random transformations.

The goal is to balance accuracy and reliability, avoiding catastrophic localization errors while maintaining good average performance.

Hosted on Kaggle · ESIT-D2I Task 1 — closed

Task 2 · Advanced Challenge

Unlearning and Data Reliability

A subset of users contributes corrupted trajectories due to faulty GPS or sensor drift. Participants will:

  • Analyze how performance degrades as corrupted data increases
  • Identify unreliable data sources once partial information is revealed
  • Develop efficient methods to remove ("unlearn") corrupted contributions without full retraining

Hosted on Kaggle · ESIT-D2I Task 2 — closed

09 · EVALUATIONScoring

High localization accuracy is crucial, but equally critical is minimizing catastrophic failures. The combined loss function balances two metrics:

MEDE — Mean Euclidean Distance Error

Average L2 distance between true and predicted positions. Reflects overall accuracy.

R90 — 90th Percentile Radius

Smallest radius enclosing 90% of predicted locations around true positions. Measures reliability and consistency.

Loss = ½ · MEDE + ½ · R90

Performance is evaluated in terms of the lowest achievable combined loss. Submissions are made through Kaggle.

10 · TEAMOrganizers

Organizing Committee

Khac-Hoang Ngo
Alejandro Lancho
Stefano Rini

Technical Support & Competition Design

Rodrigo Oliver Coimbra
Ricardo Vázquez Álvarez