Robotics

The goal was to train the autonomous robot to navigate real-world environments. The client wanted to annotate 10 thousand images obtained from the robot’s RBG camera to train the neural network. The quality of data labelling was paramount because the robot was going to navigate the streets of a metropolitan centre, and the cost of navigation error was high. It led to a large number of annotation classes

The Training Data team consisted of 70 people and was split into 5 groups, including a separate quality assurance unit. Each group was responsible for specific annotation classes and verifying other teams’ work. Teams organization in this structure allowed us to eliminate human annotation errors within our processes

The customer trained a neural network and conducted field tests that demonstrated that the robot could select the perfect route to overcome obstacles:

• 99% annotation quality
• The customer saved a lot of time delegating quality assurance to our team