Skip to main content
Back

Humanoid-GPT Trains on 2B Frames to Make Robots Move Like Humans

Humanoid-GPT trains a GPT-style Transformer on 2 billion motion frames, breaking the agility-generalization tradeoff that has bottlenecked humanoid robot control for years.

  • Humanoid-GPT is the first GPT-style causal Transformer for humanoid whole-body motion control, accepted at CVPR 2026.
  • Trained on a 2-billion-frame retargeted corpus combining AMASS, LAFAN1, MotionMillion, and PHUMA -- far larger than any prior work.
  • Uses a 3-stage pipeline: data curation, PPO-based motion expert training per cluster, then DAgger distillation into one Transformer.
  • Achieves zero-shot generalization to unseen motions (dancing, basketball, fall recovery) on a real Unitree G1 robot with no fine-tuning.
  • Establishes the first scaling law for humanoid motion tracking, showing Transformer models keep improving with more data while MLP baselines plateau.
  • Code and checkpoints are open-sourced on GitHub with deployment support for Jetson hardware and BrainCo dexterous hands.

Humanoid-GPT is a new CVPR 2026 paper that applies the same scaling recipe that transformed language models -- more data, bigger model, autoregressive architecture -- to the problem of making humanoid robots move like humans. The result is a single policy that can track highly dynamic, never-before-seen motions on a real robot, zero-shot, without any task-specific fine-tuning.

The timing matters. Humanoid robots are finally cheap and capable enough to deploy in labs and warehouses, but the software controlling their bodies has not kept up. Getting a robot to fluidly imitate arbitrary human movement -- dancing, boxing, recovering from a fall -- remains one of the hardest open problems in the field.

The wall every prior system hits

Most state-of-the-art motion trackers are built on shallow MLP (multi-layer perceptron) policies trained with reinforcement learning. They work well for the specific motions they were trained on, but they face a brutal tradeoff: prior shallow MLP trackers are constrained by scarce data and an agility-generalization trade-off. Push the model toward more agile, acrobatic motion and it loses the ability to generalize. Push it toward generalization and it becomes sluggish and conservative.

The root cause is data starvation. Most prior work trained on a few hours of curated mocap sequences from datasets like AMASS and LAFAN1. Early datasets offered high-quality but studio-constrained motions, limiting diversity. With video-based reconstruction and large-scale synthetic generation, recent datasets greatly expand motion coverage, incorporating diverse activities, styles, and subjects. But even those expansions were modest compared to what language models consume.

Scale it like a language model

Humanoid-GPT is the first GPT-style humanoid motion Transformer trained with causal attention on a billion-scale motion corpus for whole-body control. The core bet is that the same principles behind GPT -- causal (left-to-right) attention, massive pretraining data, and large model capacity -- should transfer to the physics-based control domain.