Robotics combines perception, planning and control so a physical machine can act in an uncertain world. Compared with most machine learning, the hard parts are real-time constraints, noisy sensors, physical safety, and the fact that every action changes the next observation. This section covers the core building blocks; pick a topic below.
Nearly every mobile robot, drone, or manipulator runs some version of this loop at tens to thousands of hertz. Perception turns raw sensor data into a state estimate and map (Perception & SLAM). Planning chooses where to go and how (Path Planning). Control converts the plan into motor commands that track it despite disturbances (Control Algorithms). Middleware such as ROS 2 connects these as nodes exchanging timestamped messages.
Deep learning dominates perception (object detection, segmentation, depth estimation). Reinforcement learning and imitation learning are used for locomotion and dexterous manipulation, usually trained in simulation and transferred with domain randomization. Vision-language-action models are an active research direction for general-purpose manipulation. Classical estimation and control remain the backbone where guarantees matter.
Sensor fusion, Kalman filters, visual and LiDAR odometry, and building a map while localizing in it.
PID, LQR and model predictive control: turning a desired trajectory into stable motor commands.
Graph search over occupancy grids and roadmaps to find collision-free, shortest paths.
The three topic pages map onto the three stages of the loop, and most practical problems can be traced to one of them. A robot that does not know where it is has a perception or localization problem; a robot that knows where it is but picks a poor or unsafe route has a planning problem; a robot that follows a good route badly, overshooting, oscillating, or drifting, has a control problem. Diagnosing which stage is failing comes first, because tuning a controller cannot fix a bad state estimate.
| Symptom or need | Stage | Start with |
|---|---|---|
| Position estimate drifts or jumps | Perception | Kalman filtering and sensor fusion in Perception & SLAM |
| Need a map of an unknown space | Perception | SLAM (visual or LiDAR) |
| Shortest or collision-free route on a grid | Planning | Dijkstra and A* |
| Overshoot or oscillation when tracking | Control | PID tuning |
| Tracking with constraints (speed, torque, obstacles) | Control | Model predictive control |
The stages run at different rates: control loops often run hundreds of times per second, state estimation at sensor rates, and global planning only when the goal or map changes. Keeping them as separate, well-defined components is what lets each be tested and replaced independently.