Is position always enough to predict motion?

No. For many mechanical systems, velocity and other relevant quantities also matter.

Must acceleration always be part of the state?

No. It may be computed from state and input in the chosen model; the required state variables depend on the dynamics.

Is a dynamic state the same as a sensor reading?

No. A sensor provides observations, while a state estimate may combine multiple observations and model information.

AI Glossary / Graphs and symbolic AI / Intermediate

Dynamic state

Definition

A dynamic state describes the quantities needed to model how a system will evolve under its dynamics. In the robotics context of this glossary term, it commonly combines configuration with velocity, such as a robot's joint positions and joint velocities. Position alone may be insufficient because two robots at the same location can be moving differently.

What it means

A snapshot of where a robot is does not always tell us what it can do next. A stationary cart and a cart moving rapidly through the same point have different stopping and turning possibilities. Dynamic state captures the additional motion information needed by the chosen model.
The term is model-dependent. Configuration and velocity are a standard mechanical example, but other systems may need actuator states, contact modes, or other variables. Conversely, adding every measured quantity is not automatically necessary. The state should support the prediction and control problem being modeled.

How it works

For a simple second-order mechanical model, let q denote configuration and v denote velocity. The state is x = (q,v). A dynamics model specifies how q and v change under an input such as force or torque, subject to the model's assumptions.
The controller or planner uses this state to predict motion and select actions. Measurements may not directly reveal every component, so a state estimator can combine observations with a model. A state representation intended to be Markov should contain the relevant information needed to predict future evolution from the current state and input, without depending on omitted history. Whether that approximation is adequate needs validation.

Three practical examples

A robot-arm controller considers both joint angles and joint velocities when calculating a torque command in a simulation.
A mobile-robot planner distinguishes arriving at a waypoint stopped from passing through it with nonzero speed.
A tracking system estimates a moving object's position and velocity so it can predict where the object may be after a short interval.

Think of it this way

Think of a photograph accompanied by arrows showing how the visible objects are moving. The arrows add information that the still positions alone do not provide. The analogy is limited because a useful dynamic state may include hidden variables and quantified uncertainty, and motion arrows do not replace a physical dynamics model.

Worked example

Consider a one-dimensional idealized cart. At time zero its position is p = 5 meters. Compare two possible states: state A has velocity v = 0 meters per second; state B has v = 2 meters per second. Assume constant acceleration a = -1 meter per second squared over the next second, with no other forces or constraints in the model.
Using p_next = p + v x t + (1/2) x a x t squared and v_next = v + a x t, state A evolves to p_next = 4.5 and v_next = -1. State B evolves to p_next = 6.5 and v_next = 1. The same initial position and same acceleration produce different next states because the initial velocities differ.
The negative velocity in state A simply means motion in the negative coordinate direction under this toy model. These are exact constant-acceleration calculations for an idealized system, not a real robot braking instruction. They show why configuration alone cannot distinguish the two trajectories.

Illustrative case study

At fictional Willow Robotics, a route prototype marks a waypoint reached when position enters a small region. A simulation reveals that the robot can pass through that region while still moving too quickly for the next maneuver. The team changes the modeled target to include an acceptable velocity condition.
This fictional case illustrates the difference between reaching a position and reaching a dynamic state. It does not claim that a particular tolerance guarantees physical safety or that the simplified simulator captures every real actuator and contact effect.

Common misconceptions and mistakes

Dynamic state is not simply any variable that changes over time. Its role is to summarize the system for prediction and control under a model. A position can change, yet a position-only state may omit crucial velocity information.
Velocity and acceleration are also different quantities. In a second-order model, acceleration is often determined by state and input rather than stored as an independent state component, though other models can differ. A measured state estimate is not necessarily the exact physical state. Sensor delay, noise, calibration, and unmodeled effects can all limit predictions. State completeness is therefore relative to the intended model and task.

Business use

Businesses assessing robotic or simulation systems should ask what the state representation contains and which decisions rely on it. Test cases with identical positions but different velocities, and distinguish a geometric path from a dynamically feasible trajectory. Record the timing and uncertainty of state estimates. Validate model predictions against observed motion before using them for operational commitments. A visually plausible animation does not establish that the modeled forces, velocities, and constraints agree with the real system.

Check your understanding

Two simulated carts occupy the same position, but one is stationary and the other moves forward. Are they necessarily in the same dynamic state?

Answer: No. If velocity is part of the model's state, their states differ, and applying the same input can produce different future positions.

Related terms

Configuration
Kinematic state
Dynamic model
Trajectory
velocity
dynamics model
state estimation
control input
Markov state

Learn next

Set the acceleration in the worked example to zero and calculate both positions after two seconds. Then change only the initial velocity's sign. Explain which information a position-only planner would miss in each comparison.

Sources

Tedrake, Underactuated Robotics, original author-hosted notes on mechanical state and dynamics: https://underactuated.csail.mit.edu/intro.html MIT, original state-equation teaching example using position and velocity: https://web.mit.edu/2.010/www_f00/psets/hw4_dir/tutor4_dir/tut4_b.html

Previous: Dynamic SQL Next: Dynamic time warping