INITIALIZING NEURAL CORE INICJALIZACJA RDZENIA NEURONOWEGO

overcl
RACING
cked

overclo
cked

|

We are building an autonomous AI racing driver.

Using Reinforcement Learning to achieve perfect racing lines in Formula 1.

The Vision

DRIVEN BY
EXCELLENCE

Overclocked is our team for the IBM AI racing competition project. Our goal is to build and properly train neural networks that race a formula-style open-wheel car in the TORCS simulator — from raw sensors to competitive lap times.

NEURAL_NET: ACTIVE

AERO_MODEL: BOLIDE_V1

By the numbers

What we built

0 Scored lap
0 Optuna trials
0 BC transitions
0 Model size
Politechnika Świętokrzyska

Academic Partner

POLITECHNIKA
ŚWIĘTOKRZYSKA

IBM Granite

Powered By

IBM GRANITE

Strategic Foundation

overclocked is built upon a robust collaboration between cutting-edge AI technology and academic excellence. By leveraging the advanced capabilities of the IBM Granite model, we push the boundaries of data processing and autonomous decision-making. Simultaneously, Politechnika Świętokrzyska provides us with the essential technical infrastructure and high-performance computing resources required to train and rigorously test our neural network models.

RESEARCH & DEVELOPMENT (R&D)

System Updates

Phase 6

Final agents: 92.04 s scored lap (imitation and RL)

The deliverable bc_v6.pth (429 KB) scored a 92.04 s standing-start lap under submission conditions (fuel and damag...

Phase 5

Distillation: behavioural cloning, the copycat trap, DAgger

The tuned v6 teacher was cloned into a 109k-parameter MLP (32→256→256→128→2, LayerNorm, tanh) on 500k transitions ...

Phase 4

Teacher v6: decoupled speed model + racing line

v5 computes the apex speed from the tightest forward rangefinder (v_apex = c·√sev) and adds a braking-reach term f...

Phase 4

Root cause of the 174 km/h ceiling: gear scheduling

A speed-vs-distance telemetry tool exposed a hard 174 km/h cap on every straight — invisible in lap times alone. T...

Phase 4

Teacher controllers v2–v4: structure beats tuning

Three controller families were compared. v3 (static per-position speed/line waypoints) converged to 118.8 s after ...

NEURAL STACK

Technology Dependencies Zależności Technologiczne

Click any node to learn how it fits into the project. Kliknij dowolny węzeł, aby zobaczyć jego rolę w projekcie.

Training Pipeline Ścieżka Treningu

Our learning strategy, step by step Nasza strategia uczenia, krok po kroku

Click any step to see what it means and why we chose it. Kliknij dowolny krok, aby zobaczyć co oznacza i dlaczego go wybraliśmy.

Inside the model

How the agent sees

The network never sees pixels. Every 20 ms it receives a 32-number vector: 19 rangefinder beams to the track edges plus its own speed, heading and more. Move your cursor over the track to see the beams react.

19 track-edge rangefinders · −45° to +45°

19 track rangefinders
3 speed X / Y + heading
1 track position
6 RPM, gear, 4× wheel spin
3 lap progress, time, prev-steer
32 total inputs → 2 outputs

about us

We are a team of first-year students for whom this competition became a turning point. When we signed up, we started with absolute zero knowledge of neural networks. The challenge pushed us out of our comfort zone and motivated us to spend hundreds of hours learning from scratch—from the mathematical basics of machine learning, through evolutionary strategies, to advanced Reinforcement Learning algorithms (PPO, SAC).

train_residual_sac.py — corkscrew

Meet Our Team!

Kordian Dziedzic

Team Leader, Neural Network Architect

Bartosz Guzy

Neural Network Architect

Sebastian Grzybowski

WEB architect

Janek Czajka

Web Architect

Igor Paluch

Neural Network Architect

Jakub Horążek

Social media manager

CONNECT

Instagram @hqoverclocked
X @OverclockedHQ
TikTok @hqoverclocked

Drop us a line