Here are the main areas in robotics where we can apply deep learning:

• Deep-learning-based object detector: Imagine a robot wants to pick a specific object from a group of objects. What can be the first step in solving this problem? It should identify the object first, right? We can use image processing algorithms such as segmentation and Haar training to detect an object, but the problem with those techniques is that they are not scalable and can't be used for many objects. Using deep learning algorithms, we can train a large neural network with a large dataset. It can have good accuracy and scalability compared to other methods. Datasets such as ImageNet (http://image-net.org/), which have a large collection of image datasets, can be used for training. We also get trained models that we can just use without training. We will look at an ImageNet-based image recognition ROS node in an upcoming section.
• Speech recognition: If we want to command a robot to perform some task using our voice, what will we do? Will the robot understand our language? Definitely not. But using deep learning techniques, we can build a more accurate speech recognition system compared to the existing Hidden Markov Model (HMM)-based recognizer. Companies such as Baidu (http://research.baidu.com/) and Google (http://research.google.com/pubs/SpeechProcessing.html) are trying hard to create a global speech recognition system using deep learning techniques.

• SLAM and localization: Deep learning can be used to perform SLAM and localization of mobile robots, which perform much better than conventional methods.
• Autonomous vehicles: The deep learning approach in self-driving cars is a new way of controlling the steering of vehicles using a trained network in which sensor data can be fed to the network and corresponding steering control can be obtained. This kind of network can learn by itself while driving.

Let's now look at a few deep learning libraries.