Control, Planning & Estimation

If perception is how a robot senses the world, the concepts in this section are how it acts on the world. They answer two questions that every robot has to solve, every control cycle, forever:

1. Where am I really? — sensors are noisy, biased, and incomplete, so the true state has to be estimated, not read off directly. This is state estimation.
2. What do I do about it? — given that estimate and a goal, what commands go to the motors? This is control (and the planning that decides the goal in the first place).

Tip

The one-sentence version: you cannot control what you cannot estimate, and you cannot estimate without a model. That sentence is the whole section. Estimation needs a model of how the system moves; control acts on the estimate; both lean on the same physics (dynamics) and the same math (state-space).

The loop that ties it together

Every autonomous robot runs some version of this loop, hundreds of times a second:

        ┌──────────────────────────────────────────────────────────┐
        │                                                            │
        ▼                                                            │
   ┌─────────┐    noisy      ┌────────────┐   clean    ┌─────────┐   │
   │ SENSORS │ ─ readings ─► │  ESTIMATE  │ ─ state ─► │ DECIDE  │   │
   │ (camera,│               │ filters +  │           │ plan +  │   │
   │ IMU,    │               │ state est. │           │ control │   │
   │ encoders)│              └────────────┘           └────┬────┘   │
   └─────────┘                                             │        │
        ▲                                            commands       │
        │                                                 ▼         │
        │              ┌──────────┐   torques/      ┌──────────┐    │
        └── the world ─┤ ACTUATORS │◄─ positions ───┤  (model) │────┘
            changes    └──────────┘                 └──────────┘

• Estimate turns raw, noisy sensor data into a trustworthy state. The tools here are the filters.
• Decide picks a goal (planning) and computes the commands to reach it (control).
• The model — the system’s dynamics in state-space form — is what both halves depend on. That is why a glossary section can hold estimation, planning, and control together: they share a model.

The three threads in this section

Estimation — “where am I really?” State Estimation, Kalman Filter, Complementary Filter — recover a clean state from messy sensors.

Control — “what do I command?” PID, LQR, MPC, Force / Impedance Control — turn the gap between desired and actual into actuator commands.

Planning — “where should I go?” Motion Planning — find a collision-free path for control to execute.

The model underneath Robot Dynamics, Jacobian, State-Space, System Identification — the physics and math both halves stand on.

Where to start

The entries build on each other. If you’re new to the area, read them in this order — each one assumes the ones before it:

1. Foundations — Robot Dynamics and the Jacobian (the physics: what forces move a robot, and how joint motion maps to the world).
2. The shared language — State-Space Representation (how estimation and control both describe a system).
3. Estimation — State Estimation → Kalman Filter → Complementary Filter, plus System Identification (building the model from data).
4. Control — start with PID (the workhorse), then step up to LQR and MPC (optimal and constraint-aware), and Force / Impedance Control for contact.

Note

Why this matters in practice. Modern robot-learning policies (the VLA models and imitation learning elsewhere in this glossary) still output actions that run on top of this stack — a learned policy decides what to do, but estimation and low-level control are what make the robot do it safely and accurately. The neural net is the top of the pyramid; this section is the base.

How it connects to the rest of the glossary

Perception Feeds the estimator — Sensor Fusion applies these filters across many sensors.

Fundamentals Kinematics, dynamics, and the Jacobian — the physics this section controls.

Learning from Demonstration Learned policies sit on top of this control stack.

Learning Path See where control and estimation fit in the zero-to-deployed journey.