Post-training 4B model with RL to fly Tesla Roadster 21% better than GPT-6 Astra

Researchers created a reinforcement learning environment simulating a flying Tesla Roadster and tested multiple large language models' ability to pilot it in real-time and windy conditions. A custom 4B model trained with RL outperformed all tested models including GPT-6 Astra, achieving 90% success on complex maneuvers in real time, while larger models like DeepSeek and GPT-6 Astra achieved only 0-3% due to latency issues requiring decisions every 0.2 seconds.