SignalTech/AI
Post-training 4B model with RL to fly Tesla Roadster 21% better than GPT-6 Astra
RL post-training a 4B model beats GPT-6 Astra by 21% flying a Tesla Roadster.
Context
What happened: A team built an RL environment around a physics simulator of a flying Tesla Roadster with four air thrusters, gusty wind, and the reveal's ramp flip and kickflip over a Cybertruck.
Why it matters: The results indicate that real-time control latency limits current LLMs as pilots, while RL post-training let a small 4B model outperform larger models. Everything is open source, and users can fly their own model without a GPU.
What changed
A team built an RL environment around a physics simulator of a flying Tesla Roadster with four air thrusters, gusty wind, and the reveal's ramp flip and kickflip over a Cybertruck.
Why it matters
The results indicate that real-time control latency limits current LLMs as pilots, while RL post-training let a small 4B model outperform larger models. Everything is open source, and users can fly their own model without a GPU.