Show HN: 27B open model to grade robot data 48x cheaper/27x faster vs. 6 Astra (twitter.com)

A new AI model called RoboJev has been introduced, which grades robot performance data significantly faster and cheaper than existing solutions, achieving results 48 times cheaper and 28 times faster than the 6 Astra model. Developed by Josef Chen, this model is built on the JEV-27B-VL framework, consisting of 27 billion parameters, and has been fine-tuned using over 52,000 robot episodes. Its open-weight architecture, released under the Apache-2.0 license, includes all necessary evaluation sets and associated tools, enabling broader access to its capabilities. The significance of RoboJev lies in its potential to streamline the labeling of robotic task success, a process traditionally reliant on labor-intensive human efforts. By expediting and lowering the cost of grading video data, RoboJev can enhance the efficiency of robotic training and evaluation, facilitating advancements in AI and machine learning applications in robotics. Researchers and developers can now easily test and integrate this model into their own projects, exemplified by the demo available on Hugging Face that allows users to interact with the technology using their own robot videos.