Zhongke (Shenzhen) Wireless Semiconductor and Guangxi University Win ACM MM 2026 EgoLink Challenge

A team from Zhongke (Shenzhen) Wireless Semiconductor Co., Ltd. and Guangxi University won the EgoLink Challenge at the 34th ACM International Conference on Multimedia (ACM MM 2026), held in 2026, leaderobot.com reported on September 18, 2026. According to the article, the team used brain-inspired cross-modal fusion and cognitive technology. The report said the conference was organized by the Association for Computing Machinery and initiated by Ant Group and Zhejiang University.

If leaderobot.com’s report is accurate, the focus of robot AI is shifting beyond a contest over the scale of large models toward cognitive processing that runs on the edge and the alignment of multiple sensor sources. This could favor companies with designs that integrate not only computing resources but also sensors, accelerators and development platforms.

What is new in this report is less the academic competition win itself than the presentation of a brain-inspired, cross-modal architecture designed for on-device use in robots as an integrated package that includes a chip, development platform and 4D spectral sensor. By contrast, competing in first-person-view cognition at an international conference such as ACM MM is an extension of an established trend. According to leaderobot.com, the system is described as bringing a cloud model equivalent to 100TFLOPS to the edge at 38.8TFLOPS. This suggests the focus is shifting from accuracy alone to the design of computing and cognitive infrastructure that can be installed on robots. As our related article, 36Kr reports on the issues surrounding Zidong Taichu and physical AI, showed, in physical AI the implementation of cognitive systems that run on real machines is often a greater focus than the large models themselves. This report makes that issue visible in the form of an academic competition result. The account identifies sensor-input design and the integration of inference—not just model compression—as factors behind the win. In other words, the competitive axis for robot AI may be shifting beyond the computing power of general-purpose GPUs toward how far low-latency edge processing and the alignment of multiple information sources can be advanced. The industry implications point to the importance of system-level design rather than the merits of individual components, as the report presents the 4D spectral sensor on the input side, the IDPU on the processing side and a higher-level cognitive development platform as an integrated stack. This foregrounds an implementation based on alignment at the feature or neuron layer, rather than the conventional approach of processing vision, audio and depth separately and integrating them afterward. However, the report does not establish sustained operation on real machines, computing resource consumption, the suitability of the sensor configuration for mass production or the scope of the training data. Questions to follow include how widely the 38.8TFLOPS edge requirement can be reproduced across different robots, and whether the CT-2001A and CT-HS01 can move beyond demonstration at the competition stage and be deployed externally.

For Japanese companies, the combination of edge computing circuits, sensors and development platforms described in the report could make integration of individual component strengths a key issue. However, the report does not establish how Japanese companies or Japanese components are directly involved in this case.