liveBook · Manning

Reinforcement learning

As explained in Robotics for Programmers

Reinforcement learning is machine learning in which an agent learns behavior by interacting with an environment and receiving rewards.

Reinforcement learning trains a robot, treated as an agent, to interact with an environment and receive rewards. A reward function makes desired behavior produce higher rewards, while learning seeks a policy that maps observations or environment states to actions. section 5.2

A practical robotics workflow connects learning to simulation and deployment:

  • Train the policy in simulation, where running many experiments is feasible and software failures are safer than on physical hardware. section 5.2
  • Use the learned policy to control the robot in target scenarios by converting observations or states into actions. section 5.2
  • Transfer the policy to the real robot, sometimes adapting it through a few additional real-world training iterations to address the sim-to-real gap. section 5.2 section 5.6

How Robotics for Programmers covers it

How deep it goes
Defines the agent–environment–reward–policy loop and simulation transfer, but defers learning algorithms and implementation details.
Examples
Uses simulated robots for large-scale policy training before controlling target scenarios on real robots.
What you can do after
Operate a robot with a learned policy that maps observations or states to actions in target scenarios.
How it connects
Connects machine learning with simulation, robot control, policies, observations, states, actions, rewards, and the sim-to-real gap.

Also explained in

Compare how 28 books explain reinforcement learning →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Robotics for Programmers. The section links point to the parts of the book it draws on.