> Galaxy General Robotics used 2 billion frames of motion-capture data to demonstrate, for the first time, that Scaling Laws hold in robot motion control. It's not that MLPs aren't good enough—it's that the architecture was wrong. Switch to a Transformer, and performance keeps rising with every 10x increase in data, with no saturation.
---
1. The 'Cerebellum' Dilemma of Humanoid Robots
Large models are making robots' 'brains' smarter—understanding instructions, planning tasks, holding conversations. But when you tell a robot to 'take two steps forward,' it often falls over.
Whole-Body Control (WBC) is the humanoid robot's 'cerebellum': it translates high-level commands into motor torques, coordinating 23 joints to maintain balance and track trajectories. The brain can live in the cloud; the cerebellum must run locally in real time—control loops typically require 50Hz (20ms period), meaning inference latency must stay within a few milliseconds.
Two mainstream approaches dominate today:
Classical control: MPC, WBC, LQR, etc. Stable and interpretable, but each motion requires hand-crafted models and cannot generalize to new actions.
Learned control: MLP or TCN policies trained from motion-capture data. Zero-shot generalization is their selling point, but there's a ceiling: not enough data, and the wrong architecture.
NVIDIA's SONIC pushed training to 100 million frames but still used an MLP. Experiments show MLPs start saturating around 100 million frames—adding more data no longer helps. This is the 'GPT eve' of humanoid motion control: data is piling up, but the architecture is stuck in shallow networks, so Scaling Laws never appear.
Humanoid-GPT does two things: 1. Scales data 200x—2 billion frames, unifying all major mocap datasets 2. Switches architecture—a GPT-style Transformer with causal attention
Then it asks: do Scaling Laws hold in the physical world?
The answer: yes, and beautifully so.
---
2. Dual Scaling of Data and Architecture
2.1 Data: A 2-Billion-Frame 'World of Human Motion'
Training data comes from two sources:
- Public mocap datasets: AMASS, Human3.6M, CMU Mocap, etc., all retargeted to a unified skeleton
- Large-scale in-house capture: recorded in Galaxy General's mocap studio, covering motion categories underrepresented in public data
- Causal attention: the model can only 'see' past frames, never the future—matching the physical constraints of real-time control
- Transformer decoder: stacked self-attention + MLP layers, with parameter counts from S to L
- Generative modeling: given a target motion sequence, the model autoregressively generates joint torque/position commands at each step
- Unified representation: motion targets (keypoint trajectories) and robot states (qpos, qvel) are encoded into a single token sequence
- MLP: 76.89% tracking success rate (SR)
- TCN: 81.48%
- Humanoid-GPT-S: 83.26%
- 2M frames: 88.27% SR
- 2B frames: 90.43% SR
- 2B frames: 92.58% SR
- Large TCN MPJPE: 56.15mm
- Humanoid-GPT-S (smaller model): 43.25mm
- A lead of over 30%
- GMT (one of the strongest open-source trackers)
- TWIST (industry mainstream)
- Any2Track (recent open-source method)
- SONIC (NVIDIA, 100M-frame MLP)
- TensorRT compilation: causal attention + MLP fused into efficient CUDA kernels
- C++ pipeline: no Python overhead, pure C++ inference
- Operator fusion: multi-head attention and feedforward sub-ops merged into single kernels, reducing memory traffic
- Proving Transformer + big data + causal generation works in physical motion control
- Proving Scaling Laws hold in the physical world—not language, not images, but joint torques
- Proving zero-shot generalization can go from 'individual skills' to 'an entire world of motion'
- Paper: arXiv:2606.03985 (CVPR 2026)
- Code: https://github.com/GalaxyGeneralRobotics/Humanoid-GPT
- Models: checkpoints released
- Deployment: Unitree G1, Jetson onboard, BrainCo dexterous hand support
- License: Apache 2.0
tracking/: inference core (ONNX policy wrapper, metrics, keypoint conversion)deploy/: real-robot deployment (Jetson onboard, host-side retargeting, BrainCo dexterous hand)scripts/: inference, evaluation, visualization, Gradio demoprojects/: Harmonic Motion Encoder, General Quality Selection, and other extension modules- Motion-control-first: break through hardware and low-level control; the brain can come later
- Foundation-model-first: build the general brain; classical control suffices for the cerebellum
- Paper: Qi et al. "Humanoid-GPT: Scaling Data and Structure for Zero-Shot Motion Tracking", arXiv:2606.03985, CVPR 2026
- Team: Galaxy General Robotics — Zekun Qi, Xuchuan Chen, Dairu Liu, Chenghuai Lin, Yunrui Lian, Sikai Liang, Zhikai Zhang, Yu Guan, Jilong Wang, Wenyao Zhang, Xinqiang Yu, He Wang, Li Yi
- Product: AstraBrain-WBC 0.5
- Code: https://github.com/GalaxyGeneralRobotics/Humanoid-GPT
- License: Apache 2.0
- Deployment platforms: Unitree G1, Jetson, BrainCo dexterous hand
- Baselines: SONIC (NVIDIA), TWIST, GMT, Any2Track
- Key numbers: 2B frames, 92.58% SR, 0.39ms latency, 50Hz control
Total: 2 billion frames (~200x the largest comparable research dataset). The model has seen more kinds of human motion than any single dataset could offer.
2.2 Architecture: A GPT-Style Transformer
Humanoid-GPT is not an NLP GPT transplanted onto robots—it is redesigned for motion control:
The essential difference from an MLP: an MLP is a lookup table—state in, control signal out. A Transformer 'understands'—attention captures temporal structure and whole-body coordination, learning the generative rules of motion, not a state mapping.
2.3 Experiments: Three Lines of Evidence for Scaling Laws
Evidence 1: Switching architectures alone improves performance
At the same 2 million frames of training:
Evidence 2: Transformers don't saturate with more data
Humanoid-GPT-B (fixed architecture, more data):
Humanoid-GPT-L (largest size):
MPJPE (joint position error) follows a clear power-law decrease with data volume: every 10x increase in data lowers error continuously, with no inflection point.
Evidence 3: MLPs saturate, Transformers don't
At the same 2 billion frames:
Training loss curves show diminishing returns for MLP and TCN as data grows, while the Transformer's loss keeps decreasing.
---
3. Zero-Shot Generalization: From 'Seen It to Do It' to 'Doing the Unseen'
Scaling Laws show up not only on the training set but in out-of-distribution (OOD) generalization.
Real-Robot Validation: Unitree G1
The team deployed Humanoid-GPT-B on a Unitree G1 robot, retargeting four completely unseen web dance videos to test zero-shot tracking.
Baselines:
Result: Humanoid-GPT-B achieved MPJPE lower than or equal to all baselines on all four motions.
Key point: these motions came from web videos, with no targeted data supplementation and no fine-tuning. The model generalized purely on the 'world of human motion' learned in pretraining.
Four Capabilities at Once
Prior systems usually achieve only one or two of: 1. High-DoF whole-body coordination: coordinated control of 23+ joints 2. High-dynamic motion: fast, large-amplitude actions (dancing, jumping) 3. Millisecond-level real-time response: no control-loop lag 4. High robustness: uneven terrain, external perturbations
Humanoid-GPT is the first single system to satisfy all four simultaneously.
---
4. Engineering Deployment: The Bigger Model Is Faster?
A counterintuitive result: the larger model infers faster.
| System | Avg. inference latency | |--------|------------------------| | TWIST | 2.79 ms | | GMT | ~2.5 ms | | Any2Track | ~3.0 ms | | Humanoid-GPT-B | 0.39 ms |
A 5x speedup, achieved not by being small but through dedicated kernel optimization:
The control loop holds 50Hz (20ms period); 0.39ms of inference uses only 2% of the cycle, leaving ample time for sensor reads, communication, and safety monitoring.
The lesson: scalability and inference speed are not either-or. With the right low-level optimization, a big model can be faster than a small one.
---
5. Why Call It a 'GPT-1 Moment'?
The team itself calls this 'the GPT-1 moment for the humanoid robot cerebellum.' The analogy isn't hype.
GPT-1's milestone was proving that Transformer + big data + autoregressive generation works in language, and can scale. It wasn't the strongest model, but it opened a paradigm.
Humanoid-GPT's milestone is analogous:
Previously, humanoid control learned individual skills: walking, grasping a cup, opening a door. Humanoid-GPT learns the prior distribution of human motion—how legs swing, how weight shifts, how arms and torso coordinate. Facing a new motion, it doesn't start from zero; it starts from a rich 'world model of motion.'
---
6. Comparison with SONIC: Architecture Determines the Ceiling
SONIC (NVIDIA, 2025) pushed data to 100 million frames but stuck with MLPs, which had already saturated.
Humanoid-GPT's contribution is systematically proving: the MLP bottleneck is not a lack of data, but the architecture's inherent scalability limit. Only with a Transformer can the data truly be 'digested.'
| Dimension | SONIC | Humanoid-GPT | |-----------|-------|--------------| | Data | 100M frames | 2B frames (200x) | | Architecture | MLP | Transformer (causal attention) | | Tracking SR | ~85% | 92.58% | | Zero-shot generalization | Limited | First validated OOD generalization | | Inference latency | ~2ms | 0.39ms | | Open source | No | Yes (Apache 2.0) |
SONIC proved 'data helps'; Humanoid-GPT proved 'scaling data and architecture together yields a qualitative leap.'
---
7. Open Source and Ecosystem
Humanoid-GPT is fully open-sourced:
The open-source package includes:
Any team can take the model and run it on their own robot without training from scratch.
---
8. Limitations and Industry Implications
Limitations
1. Data dependency: 2 billion frames is large, but still mocap data—not real robot interaction data. The Sim-to-Real gap remains, though real-robot validation mitigates it. 2. Dynamics vs. contact: dance and running validate high dynamics, but complex contact-rich interaction (climbing, tool use) is under-covered. 3. Causal attention information loss: not 'seeing the future' means the model can't use global motion information for planning—only local tracking. For anticipatory movements (e.g., long jumps), this may be a bottleneck.
Strategic Implications
The humanoid industry currently splits into two camps:
Humanoid-GPT offers a third path: the cerebellum itself can be a foundation model. If the cerebellum has cross-scenario zero-shot generalization, the brain and cerebellum no longer need task-specific interface plumbing—'the brain says jump, the cerebellum knows how.'
Galaxy General's AstraBrain architecture follows exactly this idea: brain (VLM) + pons (async synchronization) + cerebellum (Humanoid-GPT). The stronger the cerebellum's base capabilities, the easier the whole system scales.
---
9. One-Sentence Summary
With 2 billion mocap frames and a GPT-style Transformer, Humanoid-GPT proves for the first time that Scaling Laws hold in robot motion control—moving the humanoid 'cerebellum' from 'learning individual skills' to 'understanding an entire world of motion.' This is the GPT-1 moment of low-level robot control.
---
References