4D Vision Technology

We start with a CAD model or 3D scan. The part is trained into a digital twin environment using synthetic data. When the training reaches a point of very high reliability—99.99%—a complete vision program is provided to our customer. It is loaded on their computer running Vue software and operation of the robot can begin after a short 15-minute calibration process.

The scene is captured by pairs of 2D cameras and combined into a 3D understanding of the capture area. A single computer running Apera Vue can run up to 4 pairs of cameras to cover a large operating area.

Those images are run through our proprietary AI technology. Objects identified as pickable, even with hundreds of possibilities, are prioritized. The "most pickable" part is chosen, and full path planning instructions are provided through the robot's controller. Due to the number of simulated cycles run during the AI training process, this process happens very quickly—in as little as 0.3 seconds (3 Hz).

Conventional machine learning algorithms are based on having reference images or a CAD model as the foundation of identifying 3D positioning for objects. But what if this is a completely new part and no photos of it exist? Or the CAD model does not show how the object looks under different lighting or its texture?

In training its neural networks, Apera AI uses synthetic data. This allows us to bypass the need for reference images or relying purely on CAD models. Synthetic data allows us to improve the AI process and introduce variables such as lighting, texture, and labels or markings.