reward signal
The reward is a single number the environment hands the agent after each action, saying "that was this good, or this bad." Win a game: +1. Lose: −1. Crash the robot: a big negative number. It is deliberately simple—just a scalar—because that one number is the agent's entire definition of success. Everything the agent learns flows from chasing it.
Because reward defines the goal, designing it is one of RL's most delicate arts. Reward only what you truly want, not a convenient proxy: an agent rewarded for distance traveled may spin in circles, and one rewarded for a boat collecting points may loop a lagoon forever instead of finishing the race. Rewards can also be sparse—zero almost everywhere, nonzero only at a win—which makes learning much harder.
A reward is feedback on an outcome, not an instruction; it never tells the agent which action to take, only how good the result was.