Short answers, pulled from the story.
Apprenticeship learning is the process of a machine learning by observing an expert, rather than from explicit rules or labeled datasets. It can be viewed as a form of supervised learning where the training data consists of task executions by a demonstration teacher. It is also called learning from demonstration or imitation learning.
Inverse reinforcement learning (IRL) is a key method in apprenticeship learning that derives a reward function from observed behavior, reversing the direction of standard reinforcement learning. Instead of using rewards to teach behavior, IRL infers what goals must have motivated the expert's observed actions. It is defined by three inputs: measurements of an agent's behavior, measurements of that agent's sensory inputs, and a model of the physical environment.
AIRP was developed in 2004 by Pieter Abbeel, associated with Berkeley's EECS department, and Andrew Ng, then an Associate Professor in Stanford University's Computer Science Department. The framework addresses situations where no explicit reward function is given but an expert can be observed performing the task.
In 2017, OpenAI and DeepMind applied deep learning to cooperative inverse reinforcement learning in domains including Atari games and robot tasks such as backflips. The human's role was limited to answering queries about which of two robot actions was preferred, rather than providing full demonstrations. The researchers found evidence the techniques may be economically scalable to modern systems.
Pieter Abbeel, Adam Coates, and Andrew Ng applied AIRP to helicopter control and achieved success on demanding aerobatic maneuvers including in-place flips, in-place rolls, loops, hurricanes, and auto-rotation landings. This work was documented in a paper titled "Autonomous Helicopter Aerobatics through Apprenticeship Learning".
IRL researcher Stuart J. Russell proposed that IRL could be used to observe humans and codify their complex ethical values, with the goal of creating robots that know, for example, not to cook a cat without needing an explicit rule. He modeled this as a cooperative inverse reinforcement learning game between a person player and a robot player working to secure the person's implicit goals.