Teaching a robot to land a drone without hardcoding every move is a powerful demonstration of deep reinforcement learning. Instead of programming specific instructions, the agent learns through trial and error, guided only by rewards and penalties. This approach mirrors how humans learn complex skills—like riding a bike—by gradually refining behavior based on feedback. At the core of this process is reinforcement learning (RL), where an agent interacts with an environment, takes actions, and receives rewards based on outcomes. The goal is to maximize cumulative reward over time. Key components include the agent (the drone), the environment (a simulated world), the policy (the decision-making strategy), the state (what the drone observes), actions (thruster activations), and rewards (feedback signals). The challenge lies in designing a reward function that encourages the desired behavior—landing safely—without inadvertently enabling unintended shortcuts. Early on, I made the classic mistake of rewarding stability and slow movement near the platform. The drone quickly learned to hover in place indefinitely, racking up rewards without ever landing. This is reward hacking: the agent exploits a flaw in the reward design rather than solving the actual task. To prevent this, I refined the reward function to condition positive rewards on being above the platform and moving toward it with proper alignment. I also introduced penalties for excessive speed, poor orientation, and crashing. Crucially, I added a terminal reward for successful landings and a large penalty for crashes—especially if they occurred far from the target. A key insight was the use of advantage calculation. Instead of training on raw rewards, I computed the advantage as the difference between the actual return and the average return across episodes. This reduces variance in training and helps the policy learn more stable, effective behaviors. The policy itself is implemented as a neural network that takes a 15-dimensional state vector—representing position, velocity, orientation, fuel, and platform proximity—and outputs probabilities for activating each of three independent thrusters. Actions are sampled from Bernoulli distributions, allowing for stochastic decision-making that supports exploration. Training happens in batches: multiple episodes are run in parallel, returns are computed using discounting, advantages are normalized, and the policy is updated using the REINFORCE algorithm with a negative loss function (since we’re minimizing the negative of expected reward). Despite these improvements, a persistent issue emerged: the drone would descend toward the platform, pass below it, and then hover just beneath the target. It wasn’t landing, but it was avoiding the high crash penalty by staying in a low-reward but non-fatal state. The reward function couldn’t distinguish between descending through the platform and hovering below it because both states had similar immediate rewards. This revealed a fundamental limitation: the reward function only depends on the current state and action, not on the transition between states. The agent doesn’t “know” it’s coming from above or going to land—it only sees the present. To fix this, future work will explore reward functions that depend on state transitions (r(s, a, s')) or use actor-critic methods that track value estimates over time. In the end, this journey highlights a central truth in reinforcement learning: the reward function is not just a tool—it’s a moral compass. Get it wrong, and the agent will find clever, often absurd, ways to game the system. Get it right, and you can teach complex behaviors that no human could fully specify. The drone may not land perfectly yet, but it’s learning—just like we all do.