Reinforcement learning (RL) enables robots to learn control policies by trial-and-error interaction with an environment, maximizing cumulative reward without requiring expert demonstrations or hand-designed controllers. The robot learns a policy π(s) or π(a|s) that maps states (or state observations) to actions, optimized to maximize expected discounted reward. Core algorithms include Q-learning and policy gradient methods (actor-critic, PPO). A fundamental challenge is sample efficiency: learning in the real world is expensive (time, wear on hardware, safety risks). Solutions include (1) simulation pre-training followed by real-world fine-tuning, (2) off-policy methods like Q-learning that reuse past experience, (3) exploration strategies balancing known-good actions with discovery of better ones, and (4) reward shaping to guide learning. RL is transforming robot capabilities for manipulation, locomotion, and adaptation to new tasks, though sample efficiency and sim-to-real transfer remain critical bottlenecks.
Reinforcement learning offers a fundamentally different approach to robot control than explicit programming or behavior design. Instead of writing controllers or state machines, you specify a reward function and let the robot learn. The robot interacts with its environment (real or simulated), observes outcomes, and adjusts its policy to maximize cumulative reward. This is powerful because it can discover behaviors humans wouldn't intuitively design, adapt to new tasks quickly, and improve with more experience.
The RL Framework for Robotics: A robot perceives state s (joint angles, sensor readings), takes action a (motor commands), receives reward r (numeric signal indicating how good the outcome was), and transitions to next state s'. The goal is to learn a policy π(a|s) that maximizes expected return G = Σ γ^t r_t (discounted cumulative reward). The discount factor γ (typically 0.99) weights immediate rewards more than distant future ones. The policy can be represented as a Q-function Q(s,a) (estimated value of action a in state s) or directly as a neural network π(a|s) producing action probabilities.
Q-Learning and Deep Q-Networks (DQN): Q-learning learns the Q-function by bootstrapping: Q(s,a) ← Q(s,a) + α[r + γ max_a' Q(s',a') - Q(s,a)]. The TD error (temporal difference) r + γ max_a' Q(s',a') - Q(s,a) measures discrepancy between expected and observed returns. Deep Q-Networks scale this to high-dimensional state spaces (images, point clouds) by using a neural network to approximate Q(s,a). Experience replay stores past transitions (s,a,r,s') and samples mini-batches for updates, decorrelating samples and improving data efficiency. DQN was famously successful on Atari games; for robotics, it works for discrete action spaces but requires modifications for continuous control.
Policy Gradient Methods: Rather than estimating Q-values, policy gradient methods directly optimize the policy parameters θ by gradient ascent: ∇J(θ) ∝ E[∇log π(a|s) * (return - baseline)]. The term ∇log π(a|s) points toward actions with high log-probability; multiplying by return (or advantage, a return estimate) makes the policy more likely to repeat high-reward actions. Variants include:
Policy gradient is more natural for continuous control (commands like joint velocities or force) than Q-learning, which traditionally assumes discrete actions.
The Sample Efficiency Challenge: RL learns from trial-and-error. In simulation, a robot can collect millions of experiences cheaply. On real hardware, every interaction costs time and risks hardware damage. A manipulator learning grasping from scratch might destroy objects; a legged robot learning to walk might damage joints. Practical solutions include:
1. Pre-training in Simulation: Learn a policy in a fast, safe simulator (physics engine like PyBullet, MuJoCo), then transfer to real hardware. This saves real-world interaction.
2. Off-Policy Learning: Methods like Q-learning can learn from any past experience via importance sampling, enabling reuse of old data. On-policy methods like REINFORCE must discard data when the policy changes.
3. Exploration Strategies: Intelligent exploration (curiosity-driven learning, upper-confidence-bound exploration) finds rewarding regions faster than random exploration.
4. Reward Shaping: Adding intermediate rewards guides learning. Penalizing force in grasping tasks steers the robot toward gentle grasps without waiting for object breakage. This reduces sample complexity by orders of magnitude.
5. Demonstrations: Imitation learning (learning from human demonstrations) provides a good initial policy, then RL fine-tunes. This is more sample-efficient than learning from scratch.
The Sim-to-Real Gap: This is robotics' hardest RL problem. Simulators are abstractions; they ignore friction variations, actuator latency, sensor noise, and unmodeled dynamics. A policy optimal in simulation can fail spectacularly in the real world. Domain randomization addresses this: during training, randomize simulation parameters (friction, object sizes, dynamics) to increase distribution mismatch. This forces the policy to learn robust control, not brittle tricks exploiting simulation artifacts. If real-world parameters fall within the randomized range during training, the policy generalizes. This has enabled impressive results: robotic hands learning dexterous manipulation via large-scale simulation training and then direct transfer to hardware.
Current State: RL is transforming robotics for manipulation (grasping, insertion, dexterous control) and locomotion (walking, jumping, swimming). Sample efficiency and sim-to-real remain the limiting factors. Robots that learn in the real world are still rare for complex tasks; most successful systems combine simulation pre-training with minimal real-world adaptation. Future improvements (meta-learning, model-based RL, better simulators) will push the boundary of what's learnable in realistic time and safety budgets.
This is a foundational topic with no prerequisites.
No prerequisites — this is a starting point.