Galbot Unveils Bipedal Robot ET1 with Proprietary Embodied AI Model

On August 19, Galaxy General unveiled its bipedal humanoid robot Galbot ET1 at the WRC forum. ET1 first performed a dance routine, then interacted on stage with the robot 'Xiao Gai', which had appeared on the Spring Festival Gala. Wang He, founder and CTO of Galaxy General, explained that ET1's core lies in cross-embodiment transfer—transferring existing intelligent capabilities to different robot forms rather than training from scratch.

ET1 is equipped with the fully self-developed AstraBrain-Agent for understanding and decision-making, and AstraBrain-WBC 1.0 (the 'universal cerebellum') for mapping motion intent to full-body actions. Wang He summarized the capability as 'I See. I Learn. I Act.' In the live demo, ET1 autonomously recognized and extracted human real-time motion trajectories, converting them into its own actions, including complex street dance moves like handstands.

Wang He proposed a 'brain-pons-cerebellum' architecture: the brain handles perception, cognition, and decision-making; the cerebellum handles continuous and stable full-body movement; and the pons connects the two. AstraBrain-WBC 0.5 can already generalize to actions not seen in training, and version 1.0 further iterates. Training data includes high-precision motion capture data and action data redirected from internet human videos.

Wang He emphasized that the real challenge in entering the real world is not performing actions, but adjusting and continuing tasks when objects move, targets are occluded, or tasks are interrupted. In home scenarios, G1 completes long-horizon tasks such as organizing items, doing laundry, and preparing breakfast. During these tasks, if someone moves a bread loaf or takes a cup, the robot must re-recognize the state and adjust its actions.

In retail scenarios, G1 faces dense shelves and must identify targets in real time and adjust its picking strategy. Galaxy General's smart pharmacy solution has entered instant retail warehouses in multiple cities, and the 'Galaxy Capsule' has also been deployed. In industrial scenarios, S1 handles kilogram-level material depalletizing, transporting, and palletizing, and has entered 24/7 normal operation on CATL's production line.

These capabilities correspond to AstraBrain-WAM 0.5 (World Action Model), which models the future in latent space, reducing interference from lighting and textures. It has been used for tasks like supermarket loading/unloading, folding clothes, and moving boxes. Wang He noted that long-horizon task capabilities come from combining the WAM foundation model with real-robot reinforcement learning, extending to 10-minute continuous breakfast tasks.

Wang He introduced Galaxy General's five-layer data system, AstraData (Galaxy Star Data), which includes internet data, human action data, synthetic simulation data, real-robot teleoperation data, and real-robot feedback data. Each type serves a different purpose: internet and human video data provide semantic and operational priors; synthetic simulation data expands training scale; real-robot teleoperation data provides high-precision trajectories; and deployed real-robot data feeds back for iteration.

For motion control, the team observed that as data scale increases, joint tracking error continues to decrease, indicating that full-body motion control follows the same scaling path as foundation models. ET1 is also learning complex motor skills like tennis, which requires perception, decision-making, and motion control to be completed continuously in extremely short time frames. Wang He views this as one of the representative tasks for the 'AlphaGo moment' of humanoid robots.

The forum released the 'Global Embodied Intelligence Large Model Research Report', proposing an evaluation framework for embodied large models, covering four dimensions: data infrastructure, end-to-end capabilities, general capabilities, and deployment capabilities. It further proposes an L1-L5 tiered system to measure the development level of models from single capabilities to multi-scenario, multi-task, continuous learning, and autonomous closed-loop.

Another release, Caixin Data's 'Embodied Intelligence Long-Term Value Index', examines companies from dimensions such as intelligence boundaries, data foundations, and model architecture, focusing on how technical capabilities translate into long-term industrial value. The forum's roundtable discussion focused on the evolution path from VLA to world action models, and the gap between laboratory progress and real-world scenarios.

BAIC Penglongxing has developed 4S store reception and shopping guide applications based on the Galbot platform, with partners like NVIDIA and CATL participating in the development ecosystem. Wang He previewed that on August 22 at 7:30 PM, Galaxy General's humanoid robot will perform a mixed human-robot doubles performance at the opening ceremony of the World Humanoid Robot Games, and will interact with tennis guests.