JoyIn Introduces ‘AETHER Space’ Foundation Model for Humanoid Robots

Humanoid.Guide reported on Sept. 21, 2026, that JoyIn (Suzhou JoyIn Intelligent Technology) had introduced ‘AETHER Space,’ a foundation model for humanoid robots. The article said the model is designed to connect memory, decision-making, perception and response control in a single software stack.

By highlighting a foundation model that includes vision and physical control for humanoid robots, JoyIn underscores the importance of training-data design alongside robot hardware performance. If the company’s claim that the model reduces reliance on robot-specific data is accurate, it could affect the cost structure of data collection.

What is new about AETHER Space is its effort to combine humanoid robot perception and control around visual observations and physical context, rather than focusing on language. JoyIn was reported to have said it primarily uses human video data to reduce reliance on robot-specific data, which can be costly. This is a plan to connect a foundation model to the robot’s physical motion control, not simply to port conversational AI. Alongside Unitree’s announcement of a new foundation model, reported in Unitree announces G1+ and new foundation model, and efforts in motion recognition reported in Zhongke Wujiyuan introduces an embodied foundation model using motion optical flow, leads WorldArena consecutively, this story points to a humanoid foundation model race shifting from individual robot performance toward training methods and the scope of generalization. However, the main source for this report is secondary coverage, and the assessment relies on Humanoid.Guide’s account. Rather than drawing conclusions about AETHER Space’s capabilities, it is therefore important to examine the conditions under which its performance can be reproduced. As shown by related reports from Physical AI News, including four-legged robot shipments reach about 35,000 units in the first half and four-legged robot shipments increase in the first half, Chinese companies’ growing presence is evident in hardware shipments. AETHER Space, however, is focused not on shipment volumes but on how training data and control are integrated. If hardware mass production and foundation model development advance along separate tracks, competition cannot be measured solely by the number of robots produced. See the original source for details.

For humanoid and industrial robots in Japan, collecting data from real machines and integrating control systems can be challenging. If designs that rely primarily on human video data become more widespread, the handling of video data, control algorithms and real-world testing could become key factors in competitiveness.