Friction, mass, latency, lighting, sensor noise. Which axes move the sim2real needle, and which just burn GPU-seconds.
Not every randomization axis earns its GPU-seconds. We ablated five — friction, mass, actuator latency, lighting, and sensor noise — to see which actually move the sim2real needle on manipulation tasks.
Friction and actuator latency were the heavy hitters: randomizing them closed most of the gap on contact-rich tasks. Lighting helped perception-bound policies but did little for blind manipulation.
Sensor noise was a double-edged sword — a little improved robustness, too much drowned the signal and just burned compute. Mass randomization mattered only when the payload varied in the real task.
The takeaway: randomize the axes your real environment actually varies, and measure transfer instead of assuming it. SimArena makes the measurement cheap enough to do every time.