Leju Robotics Releases Vertical-Specific Embodied AI Model KUAVO VLA

The concept behind KUAVO VLA is a “vertical-domain model” positioned between a general base model and specific robotic skills. Leju says pre-trained VLA models often fail to fully perform in concrete settings, while post-training requires thousands of data samples and repeated collection, training and debugging. By conducting an additional round of mid-training focused on industrial scenarios, Leju embeds common capabilities of the KUAVO body and industrial environments into the model, narrowing the learning scope for new skills. This shifts development from relearning to rapid adaptation on existing abilities.

The model incorporates 100 “industrial common capabilities,” including sorting, handling, loading/unloading and assembly, letting robots complete tasks via atomic actions without extensive post-training. All real-robot data used for training came from the single KUAVO configuration, which reduces interference from different hardware platforms and improves inference accuracy on the robot.

Leju evaluated KUAVO VLA on 25 tasks: 20 custom industrial tasks and five from the public GM100 task set. It compared the model with general-purpose baselines π0.5, GR00T N1.7 and Lingbot-VLA 2.0. According to the report, KUAVO VLA achieved an overall task success rate of 48.27% and a process score of 74.51%, ranking first in both metrics. Lingbot-VLA 2.0 scored 16.27% in success rate and 42.06% in process score, meaning KUAVO VLA improved by 32 and 32.45 percentage points respectively. The company said the model also showed smoother motion, more stable decisions and fewer repeated corrections in real-robot operation.

Leju’s executive vice president Ke Zhendong told Leiphone that the company aims to support secondary developers by providing a platform, models and toolchains. Leju has run development competitions, offered robot and toolchain training, and built a dedicated team to help developers solve practical issues. Ke said the team’s key performance indicator is obtaining real user feedback on toolchains and resolving their problems. He also said that a platform that lets developers earn money is the best platform, and that KUAVO VLA is intended to lower the threshold, reduce costs and speed up skill delivery.

According to the report, Leju has accumulated large volumes of suitable mid-training data from its industry-leading training grounds and real production lines. The company does not bind itself to a specific base model, allowing developers to migrate across models through partnerships while still benefiting from performance improvements. The model reflects a broader industry shift toward serving secondary developers so that embodied intelligence can evolve through an open ecosystem.