LumiBot Targets Embodied AI Data Bottlenecks at IROS 2026

The core thesis behind LumiBot—also known as Shiyue Technology—is that the upper bound of AI performance is strictly constrained by hardware reliability. While current multimodal datasets for dexterous hands often total only a few tens of thousands of samples, training robust models requires millions of hours of operational data. To bridge this gap, the company has developed a full-stack ecosystem including the Lumi-Dex hand series and Lumi-Tac visuo-tactile sensors.

Technological integration remains the company's primary focus. The Lumi-Dex-21 hand utilizes a drive-control design that combines linkage and direct-drive mechanisms, while its upcoming flagship model supports over 5,000 hours of continuous operation. Complementing this, the Lumi-Tac V2 sensor provides 10-micron spatial resolution and a 120 fps frame rate. These components feed into the Lumi-Mind foundation model, which uses a Mixture of Transforms architecture to fuse vision, tactile, language, and action inputs.

During live demonstrations at IROS, the firm showcased the Lumi-Dex-11 hand executing complex, long-horizon tasks—such as stamping and tray management—while autonomously adjusting to human-introduced disturbances in real time. By moving beyond laboratory settings, the startup is currently conducting proof-of-concept validations with industry leaders in automotive manufacturing, 3C electronics, and life sciences. This strategy aims to shift the industry from small-scale research experiments toward real-world industrial deployment.