Technology
What is reinforcement learning?
A dog learns a trick from treats; a program can learn a game from its score. Reinforcement learning is the branch of AI that learns by trying, failing and being rewarded. It has beaten a Go champion, and it is how some robots learn to move. It is also the last thing we would let near a safety rule.
In short. In reinforcement learning, an agent tries actions in an environment, receives rewards, and learns which actions pay off in the long run. In 2015, one agent learned 49 Atari games from the screen’s pixels and the score alone, reaching a level comparable to a professional games tester. In 2016, AlphaGo combined it with learning from human games and beat the European Go champion 5 games to 0. Trial and error is slow and risky on real machines, so robots mostly learn in simulation, and learning within safety limits is a field of its own.
Learning from rewards
The setting has a few parts. An agent, the learner, acts in an environment: a game, a simulated room, a real robot’s world. After each action it sees what changed and receives a reward, which may be positive, negative or nothing. Its goal is not the next reward but the most reward over time, so it has to learn that a small sacrifice now can pay off later. And it has to balance trying new things, to find out what works, with doing what already works.
The idea has deep roots in psychology and neuroscience: the 2015 Atari paper notes parallels between the signals of dopamine neurons in the brain and a family of reinforcement learning algorithms. More on what dopamine does and does not do: does a dopamine detox work?
From Atari to Go
In 2015, Mnih and colleagues described a deep Q-network: an agent that received only the pixels and the game score as input. Across 49 classic Atari 2600 games, using the same algorithm, network and settings for all of them, it surpassed the earlier algorithms and reached a level comparable to that of a professional human games tester.
A year later, Silver and colleagues introduced AlphaGo. Its neural networks were trained first on games played by human experts and then by reinforcement learning from games against itself, and combined with a search algorithm. It won 99.8% of its games against other Go programs and beat the human European Go champion 5 games to 0, which the authors describe as the first time a computer program had defeated a human professional player at full-sized Go.
Why robots learn in simulation
Learning by trial and error takes a vast number of tries, and on a real robot every try takes real time and some end in a crash. So robots mostly learn in a simulator, where a fall costs nothing and thousands of tries run in parallel, and the result is then tested on hardware: sim-to-real.
Rewards can teach the wrong lesson
An agent pursues exactly the reward it is given, not what its designers meant. A badly chosen reward can teach a shortcut nobody wanted, and a behaviour learned in one setting can fail in another that looks similar. Learned behaviour therefore needs testing in many more situations than it was trained in.
Safety is a field of its own
García and Fernández’s 2015 survey of this field describes it as learning to maximise reward in problems where it matters to keep reasonable performance or respect safety constraints, while learning, while in use, or both. They group the approaches in two: changing what counts as the best behaviour, by adding a safety factor, and changing how the agent explores.
How we use learning
Alkyon, our concept for a flying home companion, is designed so that learned parts only propose: where to look, which shape to fold into, how to approach a landing spot. Tested, fixed code checks every proposal, and the safety guard, which keeps 1 m from people, limits the speed and stops the motors on contact, does not learn. Learning motion strategies in simulation is a later step in our plan, not something Alkyon does today: AI proposes, code decides. It is a concept in development, for Europe only.
Sources, read 4 November 2026: Mnih and colleagues, human-level control through deep reinforcement learning, Nature, 2015; Silver and colleagues, mastering the game of Go with deep neural networks and tree search, Nature, 2016; García and Fernández, their survey of reinforcement learning under safety constraints, Journal of Machine Learning Research, 2015. General information.
Questions
What is reinforcement learning in simple terms?
A way for a program to learn by trial and error: it acts, receives a reward or a penalty, and gradually prefers the actions that earn the most reward over time.
Is reinforcement learning used in robots?
Increasingly, and mostly in simulation first, because trial and error on real hardware is slow and can break things. Learning while respecting safety limits is a research field of its own.