China Telecom (TeleAI)

“a policy trained entirely on sim-generated data achieves comparable or superior zero-shot performance on a real Unitree G1 humanoid compared to a policy trained on real-robot teleoperation data”

“The policy trained on simulation data alone achieves a real-robot success rate comparable to, and on some tasks higher than, the one trained on real-robot data... We attribute the cases where simulation even surpasses real data to visual diversity”