Interview with Jisu Tech Co‑Founders: Pioneering a Data‑Quality ‘Refinery’ for Spatial Intelligence

Jisu Tech’s three co‑founders discuss the massive data‑collection wave for embodied intelligence and why sheer volume cannot solve the industry’s chronic data‑quality problem. They argue that quality, not quantity, is the bottleneck for training reliable models.

Three Pillars: A New Quality Paradigm

One Map : Precise spatial mapping using surveying techniques to predict and evaluate factors that affect data quality.

One Ruler : An error‑detection model that measures scene‑specific measurement errors and suggests corrections.

One Flywheel : Accumulated expert judgments from diverse sensors and scenarios are encoded as parameters of a quality model, closing the loop for downstream delivery.

These pillars form Jisu Tech’s technical moat in the data‑quality niche.

Quality Factor Library (QFLab)

The company has built a multimodal factor library containing over a billion frames of data, distilled into thousands of high‑quality factors. Each factor quantifies how objects (e.g., glass, leaves, reflective surfaces) impact sensor accuracy, providing both qualitative and quantitative assessments for different modalities such as LiDAR, vision, inertial navigation, and GNSS.

Training the library involves aligning observations with ground‑truth measurements, a process inherited from traditional surveying where error quantification is paramount.

From NASA Lessons to Autonomous Driving

Co‑founder Zhou Yuan cites his work on NASA’s Mars rovers, where precise landmark weighting reduced positioning error from tens of meters to a single meter, extending mission life dramatically. This experience shaped Jisu Tech’s belief that data quality can be treated as a separate modality, comparable to vision or language.

The company’s quality model, slated for full training by the end of 2025, already outperforms human inspectors in most re‑checks, and its factor library now covers more than 5,000 km of driving scenarios.

Commercial Path and Impact

Jisu Tech delivers high‑precision spatial datasets (e.g., 7 surround cameras, 1 LiDAR, 1 GNSS per vehicle) with sub‑centimeter error, 100 % feature completeness, and >99 % classification accuracy. A 2024 contract for 395 000 000 data groups was fulfilled in 35 days, leading to repeat orders and a conversion rate above 80 %.

By re‑using over 60 % of historical data, the company reduces collection cost and accelerates model iteration, turning raw data into a sustainable, reusable asset.

Future Outlook

Jisu Tech sees the data‑quality “flywheel” as the fastest way to scale embodied AI, especially in household robotics where high‑frequency human‑machine interaction generates valuable training signals.

Ultimately, the firm treats spatial data like crude oil: raw data must be refined through rigorous measurement to become the backbone of the intelligent‑era infrastructure.