Students from the AI Talent Hub at ITMO University have developed a prototype service named Praxis for IT Imperial. The tool is designed to automate the labeling of human action videos used for training robots. By generating a preliminary draft, the system aims to reduce manual labor in the data preparation process by at least 80 percent.
The current manual processing of a short video clip takes between 15 and 30 minutes. In internal tests, the prototype generated a draft for a 30-second video in approximately one minute, with an additional three minutes required for human verification. Experiments showed that using the automated draft makes the labeling process 2.2 times faster compared to working from scratch.
The system utilizes a dual-model architecture: one model identifies action boundaries, while the other classifies the specific action and object. Testing on 85 clips revealed an average boundary error of 0.27 seconds, with 90 percent of boundaries falling within a two-second tolerance. In a separate end-to-end test, 72 percent of action steps matched the ground truth.
Despite these results, the developers noted that content recognition and data transfer capabilities require further refinement. The team will continue collaborating with IT Imperial throughout the semester to test the prototype on corporate data and measure performance metrics. IT Imperial plans to integrate these findings into their internal processes.
