In a significant development, Galaxy General Robotics, a company specializing in embodied intelligence, has officially launched the world's first humanoid robot general cerebellum GPT foundation model, AstraBrain-WBC 0.5.

This model is a key component of the "AstraBrain" technological framework and marks the first time GPT-style scaling laws have been introduced into the field of robot motion control, achieving several world-class breakthroughs.

It has established the world's largest humanoid robot motion corpus, validated scaling laws similar to GPT in motion control tasks, and for the first time on a physical robot, realized zero-shot generalization for executing a large number of unseen actions.

This endows robots with motion intelligence to handle unfamiliar tasks.

From Learning Single Actions to Understanding Motion Logic

A long-standing divide in embodied intelligence exists: if the "brain" is responsible for perceiving and understanding the world, the "cerebellum" determines a robot's ability to act in the physical world.

This involves coordinating dozens of degrees of freedom across the body within milliseconds, maintaining balance amidst complex dynamics, resisting disturbances, and executing precise movements.

However, traditional motion control heavily relies on specialized training for individual skills, requiring redesign or fine-tuning for every new action or scenario, consistently failing to break the generalization bottleneck.

AstraBrain-WBC 0.5 is a direct response to this challenge. It proves for the first time that robot motion control also follows the scaling laws of large models.

By continuously expanding data scale and model capacity, the model learns the underlying universal logic of motion rather than merely memorizing specific actions.

Consequently, robots can achieve zero-shot generalization for executing new actions and scenarios never encountered during training.

Training with 20,000 Hours of Human Motion Data to Create the Industry's Largest Humanoid Robot Motion Corpus

The capability of large models stems from scale, a principle validated for the first time with AstraBrain-WBC 0.5.

Galaxy General's intelligent data foundation, "AstraData," innovatively encompasses internet data, human behavior data, simulated synthetic data, real robot teleoperation data, and real-world scenario feedback data, forming a complete closed loop from data production and model training to scenario validation and iterative optimization.

The 20,000 hours of motion data cover diverse scenarios including dance, sports, collaborative handling, and industrial operations, with the range of motion space being 4 to 5 times larger than the widely used AMASS dataset.

Building on this, the research team further scaled the model to 80.4 million parameters, making AstraBrain-WBC 0.5 the world's first humanoid robot whole-body real-time motion control large model at the scale of GPT-1.

First Introduction of GPT-Style Architecture into Robot Motion Control, Disrupting the Long-Standing Three-Layer MLP Framework

Humanoid robot motion control models have long relied on shallow MLP networks, which, while simple in structure, have limited capacity and struggle to evolve continuously with increasing data scale.

AstraBrain-WBC 0.5 fundamentally changes this landscape. The team employed a GPT-style causal Transformer architecture for the first time, reframing whole-body control as a continuous sequence prediction problem.

This enables robots to predict future trends based on action history, achieving an understanding of "motion semantics" akin to GPT's understanding of language sequences.

Simultaneously, the research team constructed a motion prior library comprising 384 action experts and integrated them into a unified control model through distillation training, achieving a leap from a "collection of expert skills" to a "general motion foundation model."

Importantly, AstraBrain-WBC 0.5 is not simply about making the model larger. This work first validated that the field of robot motion control also exhibits development patterns similar to GPT.

As data scale expanded from millions to 2 billion frames and model size grew continuously, model performance improved steadily, and zero-shot generalization capabilities strengthened without encountering the typical performance bottlenecks of traditional motion control models.

Whole-Body Coordination, Millisecond Response, Motion Capability Approaching Human Levels

In real robot testing, AstraBrain-WBC 0.5 achieved zero-shot execution of highly dynamic actions not present in the training set, such as basketball, boxing, dancing, getting up from lying down, and collaborative handling, enabling rapid transfer without retraining.

Concurrently, the model demonstrated unprecedented robustness, maintaining stable control during rapid motion, center of gravity changes, complex contact transitions, and under strong external disturbances.

After engineering optimization, AstraBrain-WBC 0.5 achieved inference latency below 1.5 milliseconds on a single RTX 4090 and full motion capture pipeline latency under 20 milliseconds, meeting the requirements for 50Hz real-time closed-loop control.

On a 29-degree-of-freedom robot, whole-body and full-hand coordinated control was smooth and natural, with complex actions like limb coordination, center of gravity shifts, and body coordination executed seamlessly.

Accelerating Industrial Application and Establishing an Innovation Hub for Embodied Intelligence

AstraBrain-WBC 0.5 represents not just a technological breakthrough but also a key to unlocking industrial application potential.

As a motion control foundation, it can provide research institutions and developers with high-quality motion data generation capabilities, significantly lowering the barrier to training whole-body control models for humanoid robots.

Based on real-time motion tracking, developers can quickly generate creative content for dance, performances, interactive displays, and more, achieving "think it, have it."

In scenarios like emergency rescue, hazardous environment handling, and disaster site search and rescue, robots can be the first to enter high-risk areas for detection and rescue missions.

The development of Galaxy General's AstraBrain-WBC 0.5 is supported by long-term strategic planning and precise support for the artificial intelligence and embodied intelligence industries in the region.

In recent years, the area has consistently led the nation in embodied intelligence, being the first to introduce a national three-year action plan specifically for the field and building the country's most comprehensive embodied intelligence industrial ecosystem.

Data shows the area has gathered over 2,000 AI-related enterprises across the value chain and more than 300 embodied intelligence robot companies, forming a complete industrial chain centered on "brain, cerebellum, and body," successfully selected as a 2024 Ministry of Industry and Information Technology SME characteristic industrial cluster for the robotics industry.

Today, the world's first humanoid robot general cerebellum originates from here, not only heralding the dawn of the robot motion control foundation model era but also showcasing to the world the region's leading strength in hard technology sectors.

Currently, the research paper, code, and technical achievements for AstraBrain-WBC 0.5 have been fully open-sourced and made available to the ecosystem.

Galaxy General invites global research institutions, universities, developers, and industry partners to participate, jointly advancing the development of robot motion foundation models and accelerating technological innovation and industrial application of embodied intelligence.

The GPT era for the robot "cerebellum" has arrived, and this region is at the very core of this transformation.