Computer Vision for Robots: How a Robot Sees

Short answer: a robot "sees" by turning camera images into numbers and running software that finds objects, estimates distance and tracks movement. See computer vision in the Atlas.

What vision lets a robot do

• Detect objects: find a cup, a person or a pallet (object detection).
• Estimate position: work out where the object is so an arm can reach it.
• Follow lines or lanes: simple colour or edge detection is enough.
• Help with mapping: camera-based SLAM.

A simple example

A sorting robot must separate red and green apples. A basic program thresholds colour and drives a motor. No neural network is needed. For mixed, tilted or dirty items, a trained model works better.

Common pitfalls

• Lighting changes. Glare or darkness can break a model trained in a bright room. Test in the real lighting.
• Not enough varied training images. The robot works on the test table and fails elsewhere.
• No distance information. One camera does not directly give depth. Add stereo cameras, lidar or a depth camera. Compare them in LiDAR vs radar vs camera.
• Slow processing. Large models can lag on small boards, so a delayed answer is a wrong answer for a moving robot. See edge AI.

Getting started

1. Use a USB camera with a Raspberry Pi and try colour tracking.
2. Move on to a pretrained detector.
3. Measure how fast it runs on your board before building the robot around it.

FAQ

Do I need a GPU? Not for simple vision. For modern detection models, an accelerator helps a lot.

Camera or lidar for a first robot? Start with a camera, because it is cheap and teaches the most.

Overview: AI and robotics, the complete guide.