GoKawiilRL-trained 4B model beats GPT-6 Astra in simulated flying Roadster test
A research team built a physics-based RL simulation of Tesla's flying Roadster, complete with wind gusts and the ramp-flip and Cybertruck kickflip stunts from the reveal, then had several leading LLMs try to pilot it. Models including DeepSeek V4.1 Flash, Nemotron 3.5 Super, Muse Spark Open, Beam 1.0, and GPT-6 Astra mostly crashed, landing few or none of the maneuvers even with unlimited thinking time, while a small 4B model fine-tuned with reinforcement learning landed reliably, including in a 33 m/s storm.
GoKawiil's interpretation of the reporting above, not reported fact.
The result suggests that raw language-model scale and reasoning time don't translate well into fast, continuous control tasks that require a new decision every 0.2 seconds, an area where specialized reinforcement learning may still outperform general-purpose LLMs. It could serve as evidence, as the researchers imply, that purpose-built RL post-training remains more effective than prompting large general models for real-time robotic or vehicle control problems.
- A 4B parameter model trained with reinforcement learning outperformed much larger LLMs, including GPT-6 Astra, in a simulated flying Roadster test.
- Leading open models like DeepSeek V4.1 Flash and Nemotron 3.5 Super took 1–6 seconds per decision, far too slow for a control loop needing actions every 0.2 seconds.
- Even when given unlimited time to think, these large models still crashed more often than the small RL-trained model, especially in high-wind conditions.
Source: skypilot.ai, 2026-10-30
Published there as: “Post-training 4B model with RL to fly Tesla Roadster 21% better than GPT-6 Astra”
Read the original report → The summary and analysis above are GoKawiil's own, written from reporting by the source above. Facts and quotes belong to the original publisher.