The capital will fuel the expansion of a specialized capture network designed to record skilled workers performing manual tasks—ranging from folding laundry and loading dishwashers to complex warehouse logistics. Unlike traditional crowd-sourced annotation, Realset employs domain experts to generate data in real-world environments. This approach produces synchronized video, IMU, and action logs specifically optimized for vision-language-action training models.

Beyond data acquisition, the company is developing the Realset Workspace, a platform supporting operations in six languages, including English, Japanese, and Simplified Chinese. A key component of their roadmap involves launching open benchmarks to test how AI policies perform outside the lab. The company plans to release its first benchmark focused on household manipulation by Q4 2026, targeting models such as OpenVLA and GR00T. Founder Hunter Guo notes that the current bottleneck in AI development is no longer labeling, but the capture and verification of physical human expertise.