Zekun Qi

“Humanoid-GPT trains on a 2-billion-frame corpus, representing a 200x scale-up... achieves a 92.58% tracking success rate across unseen motions.”

“their largest model (Humanoid-GPT-L, 80.4M parameters, trained on 2B frames) achieves a 92.58% tracking success rate across unseen motions, compared to MLP baselines trained on 6–9M frames that hover around 76–81% (Table 2). This is a 16+ point gap in real deployment stability.”

“Our tracker reproduces these motions in real time without any task-specific fine-tuning, demonstrating strong zero-shot transfer from simulation to the real world.”

“Zekun Qi — Tsinghua University. Co-author, also appears on SoFar (language-grounded spatial reasoning for manipulation) and ShapeLLM (3D object understanding for embodied interaction).”