Team Sqro — IBM AI Racing League

AI racing driver for TORCS Corkscrew track using Behavioral Cloning.

How it works

• collect — rule-based driver collects 30 laps of training data
• bc — neural network trained via Behavioral Cloning (steer + accel + brake)
• play — AI drives autonomously, code handles gears only

Architecture

• Input: 14 features (angle, track position, speed, 9 track sensors)
• Hidden layers: 256 → 128 → 64 neurons (ReLU)
• Output: steer (tanh), accel (sigmoid), brake (sigmoid)
• Training: 500 epochs, Adam optimizer, MSE loss = 0.0013

Usage

• python train.py collect # Phase 1: collect training data
• python train.py bc # Phase 2: Behavioral Cloning
• python train.py play # Run trained AI driver

Results

Lap time: ~1:42 from standing start on Corkscrew track

Team

Team Sqro — Silesian University of Technology
IBM AI Racing League 2026