Humanoid-GPT: A GPT Moment for Humanoid Robot Locomotion Control

Forum topic · 小凯 · 2026-06-20

Summary

GalaxyGeneralRobotics has released Humanoid-GPT, a GPT-style Transformer for humanoid whole-body control, and claims the first validation of a Scaling Law in robot motion control. Trained on 2 billion frames of unified motion-capture data (AMASS, Human3.6M, CMU Mocap, plus proprietary captures) — roughly 200x more than prior work — the model achieves 92.58% tracking success rate at its largest size, versus 76.89% for an MLP baseline. Unlike MLPs, which saturate around 100M frames (as seen in NVIDIA's SONIC), the Transformer keeps improving as data scales 10x. Deployed on a Unitree G1 robot, it zero-shot tracks dance motions retargeted from unseen web videos, matching or beating TWIST, GMT, Any2Track, and SONIC. Via TensorRT compilation, operator fusion, and a pure C++ pipeline, inference latency drops to 0.39ms — faster than smaller models — fitting easily in a 50Hz control loop. The system combines high-DOF whole-body coordination, dynamic motion, real-time response, and robustness in one model, and is open-sourced under Apache 2.0 (arXiv:2606.03985, CVPR 2026).

GalaxyGeneralRobotics' Humanoid-GPT demonstrates that Scaling Laws hold for humanoid robot motion control — provided you replace the field's default MLP architecture with a GPT-style causal Transformer and scale data 200x beyond prior work.

The "cerebellum" problem

Large models make robot "brains" smarter, but whole-body control (WBC) — translating high-level commands into torques for 23+ joints while balancing in real time — remains a bottleneck. Control loops run at 50Hz (20ms cycles), leaving only a few milliseconds for inference. Classic methods (MPC, WBC, LQR) are stable but don't generalize; learned MLP/TCN trackers hit a data ceiling — NVIDIA's SONIC saturated around 100M frames.

Key points

Data: 2 billion frames (~200x prior work), unifying AMASS, Human3.6M, CMU Mocap and proprietary mocap onto one skeleton.
Architecture: A Transformer decoder with causal attention, autoregressively generating joint commands from tokenized motion targets and robot state — learning the *generative structure* of motion rather than a state-to-control lookup.
Scaling evidence:
At 2M frames: MLP 76.89% tracking SR, TCN 81.48%, Humanoid-GPT-S 83.26%.
Humanoid-GPT-B: 88.27% SR at 2M frames → 90.43% at 2B frames; Humanoid-GPT-L reaches 92.58% SR.
MPJPE follows a clear power-law decrease with data; MLP/TCN losses plateau while the Transformer keeps improving. At 2B frames, even Humanoid-GPT-S (43.25mm) beats a larger TCN (56.15mm) by over 30%.
Zero-shot generalization: On a Unitree G1, Humanoid-GPT-B tracks four unseen dance videos retargeted from the web, matching or beating GMT, TWIST, Any2Track, and SONIC — no fine-tuning, no task-specific data. It is claimed to be the first single system combining high-DOF whole-body coordination, high-dynamic motion, millisecond real-time response, and robustness.
Counterintuitive speed: Despite being a larger model, inference runs at 0.39ms vs ~2.5–3ms for competitors (TWIST 2.79ms, GMT ~2.5ms), via TensorRT-compiled fused kernels and a pure C++ pipeline — only 2% of the 20ms control cycle.

Why a "GPT-1 moment"

Like GPT-1, the significance is not being the strongest model but opening a paradigm: Transformer + big data + autoregressive generation works for physical control, and it scales. The robot learns a prior over the human motion world rather than individual skills, enabling zero-shot generalization to new actions.

Limitations

Training data is mocap, not real robot interaction; Sim-to-Real gap persists (mitigated by real-robot tests).
Complex contact interactions (climbing, tool use) remain underexplored.
Causal attention cannot plan with future information, potentially limiting proactive movements.

Open source

Released under Apache 2.0 (paper: arXiv:2606.03985, CVPR 2026; code: https://github.com/GalaxyGeneralRobotics/Humanoid-GPT), with checkpoints, deployment tools for Unitree G1 / Jetson / BrainCo dexterous hand, evaluation scripts, and extension modules. The work underpins GalaxyGeneralRobotics' AstraBrain architecture: VLM brain + bridge + Humanoid-GPT cerebellum.