Efficiency Over Scale: Why Small RL Models Are Winning the Robotics Race

In the current arms race of artificial intelligence, parameter count has long been viewed as the primary proxy for capability. However, recent testing by SkyPilot suggests that for physical, real-time robotics, raw scale is becoming a vanity metric. When the objective shifts from generating text to executing agentic maneuvers in a physics-driven environment, specialized efficiency outweighs sheer size.

To test this, as SkyPilot first reported, the team built an RL environment around a physics simulator of a flying Tesla Roadster—inspired by a recent demonstration where the vehicle performed a ramp flip and a kickflip over a Cybertruck. The simulation, which runs at 500 Hz, requires the pilot to make decisions every 0.2 seconds to manage four air thrusters and navigate varying wind conditions, ranging from light to storm levels (up to 33 m/s).

SkyPilot pitted some of the industry's most prominent large-scale models against this environment, including OpenAI's GPT-6 Astra, NVIDIA's Nemotron 3.5 Super VL, Meta's Muse Spark Open, Reflection's Beam 1.0, and DeepSeek V4.1 Flash. The results highlight a critical failure in latency. In real-time trials, these models took between 1 and 6 seconds to provide an answer, whereas the car requires a new action every 0.2 seconds. Consequently, GPT-6 Astra and the other open models landed 0% of the kickflip stunts on the clock.

Even when the "world was paused" to remove the latency penalty, the massive models struggled to outperform a much smaller, specialized alternative. While DeepSeek V4.1 Flash achieved an 87% success rate with unlimited time to think, and GPT-6 Astra landed 74%, they were still outperformed by a 4B model trained with reinforcement learning (RL).

According to SkyPilot, the RL-trained Qwen3-4B—which cost approximately $13 to train—became the superior pilot. In real-time scenarios, the Qwen3-4B answered in 0.09 seconds and successfully landed 90% of kickflips.

This disparity underscores a fundamental shift in the requirements for agentic AI. The flying Roadster simulation requires precise, rapid-fire adjustments to telemetry—such as managing roll and pitch while upside down—where a delay of even a few seconds results in a crash. The fact that a 4B model can outpace the industry's largest LLMs suggests that the future of robotics does not lie in larger models, but in RL-tuned, smaller architectures optimized for speed and specialized physical interaction.