Galbot (Galactic General Robotics)

Galbot (Chinese name 银河通用机器人) is a robotics company founded in 2023 as a general-purpose robot company built around an embodied foundation model — a single learned model meant to let a robot perceive, decide and move, rather than a separate hand-written program for every task ^[1].

Where it comes from

Galbot was set up in 2023 as a general-purpose robot company built around an embodied foundation model ^[1]. That framing shapes how the company presents its work: the hardware is one part of the story, and the training data, the models and the places where the robots actually work are treated as parts of one system rather than separate products.

The AstraBrain models

The centre of this work is AstraBrain, which Galbot presents as the first end-to-end embodied model that integrates the brain, the cerebellum and neural control in a single model ^[1]. In rough terms, the "brain" layer is the high-level part that decides what should be done; the "cerebellum" layer turns a decision into coordinated whole-body movement; and neural control is the low level that drives the joints. Building them as one end-to-end model means the layers are learned together instead of being designed separately and stitched together afterwards.

AstraBrain-WBC is described as the company's general cerebellum model, the part responsible for whole-body control. It has 80.4 million parameters and was trained on 20,000 hours of human motion data ^[1]. Human movement is a practical source of training material for this kind of model: people produce a far wider range of motion than one robot can collect by itself, so recorded human motion gives the model many examples of how a body moves.

Galbot also reports how the model behaves as the training set gets larger. When training data grew from 2 million to 2 billion frames, it says the success rate of whole-body control rose from 83.26% to 92.58%, and that on real robots the model generalised zero-shot to motions it had not seen during training ^[1]. Zero-shot here means the robot carries out an unfamiliar movement without being retrained for that specific task.

Data infrastructure

Behind the models sits AstraData, a data infrastructure the company describes as being at the hundred-billion scale ^[1]. For embodied AI, data is usually the binding constraint: a model that has to handle the physical world needs large amounts of movement and interaction to learn from, and the pipeline that collects, cleans and organises it matters as much as the model architecture.

Robots and deployments

Galbot's products include the Galbot G1 and the Galbot S1, which the company says are deployed in retail, industry, logistics, healthcare and services ^[1]. In retail it gives a more specific figure: its robots run in unmanned retail cabins at more than 170 sites in over 40 cities ^[2]. Unmanned cabins are a comparatively contained setting — shelves, a known layout, and a steady stream of picking tasks — which makes them a useful place to accumulate real operating experience.

The tennis demonstration

Galbot says its robots achieved the first autonomous humanoid tennis rally, using a whole-body real-time planning and control algorithm ^[1]. A rally is a demanding test for a humanoid: the ball keeps moving, the robot has to reposition its whole body rather than a single arm, and the response has to happen in real time, so perception, planning and control are exercised in one loop.

Where it sits

The company says it is the only firm to close the loop from data infrastructure to models and to real-world deployments ^[1] — a claim about its own position rather than an independently measured result. Its next step in that direction is a native agent robot called 银河星仔 (Galbot Star), on which it opened pre-orders on 3 September 2026 ^[3].