Does the robot physically feel the field?

Not necessarily; the controller computes it.

Will increasing repulsion always help?

No. The example shows that a larger gain can cancel progress at a particular position.

Does a potential field require machine learning?

No. Its functions and parameters can be specified directly.

AI Glossary / Graphs and symbolic AI / Advanced

Potential field

Definition

A potential field is a scalar function over positions or robot configurations that guides motion through its slope. In artificial potential-field planning, goals commonly create attraction and obstacles create repulsion. A controller uses the negative gradient to move toward lower potential. The field is mathematical, not necessarily physical.

What it means

Instead of first finding a complete route, a simple field-based controller repeatedly asks which nearby direction decreases its potential. The goal influences that direction, and nearby obstacles modify it. This can produce responsive local motion, but local guidance alone does not guarantee that the robot will reach the destination. The shape of the field matters: an undesirable stationary point can stop motion, and a poor geometric model can overlook collisions involving the robot's body.

How it works

Define an attractive potential toward the target and repulsive terms near obstacles, then add them. For a simple point model with goal g, the attractive term can be one half of the squared distance to g. Its negative gradient points toward g.
A repulsive term can depend on distance to an obstacle and vanish outside a chosen influence range. The summed negative gradient becomes a desired direction, velocity or force, depending on the controller. These are different interpretations: an actual force-driven robot has dynamics, and damping may be needed to settle. The field can be defined in configuration space, which accounts for robot configurations, or constructed through selected points on the robot. Distance calculations and the mapping into feasible motion require appropriate geometry.

Three practical examples

A simulated warehouse vehicle combines attraction toward a waypoint with repulsion from nearby obstacles. A route planner can change the waypoint when local guidance fails to make progress.
A robot arm uses repulsive effects at points along its links, not just at its tool tip. Avoiding an obstacle with the tip alone would not establish that the rest of the arm avoids it.
A mobile-robot experiment updates obstacle-related terms as its environment changes. The experiment measures stopping, oscillation and clearance as well as success in reaching the goal; fast field evaluation alone is not proof of reliable navigation.

Think of it this way

Picture an artificial landscape with a valley near the destination and hills around obstacles. Moving downhill offers a convenient local instruction. However, a smaller basin or a perfectly balanced slope can prevent progress toward the intended valley. This metaphor supplies no route guarantee.

Worked example

Consider a point robot in a flat mathematical workspace. Its position is q = (1, 1), its goal is g = (4, 1), and a point obstacle is at o = (2, 1). Use the attractive potential Uatt = 0.5 times ||q - g||^2.
Let d be the distance from q to o. For d below 2, choose Urep = (eta/2) times (1/d - 1/2)^2; beyond that range, set it to zero. This example stays away from d = 0, where the formula is undefined. Set eta = 2.
At the starting point, d = 1. The attractive negative gradient is (3, 0). The repulsive negative gradient points directly away from the obstacle and has magnitude eta times (1/d - 1/2) divided by d^2. Its value here is 2 times 0.5 = 1, giving (-1, 0). The sum is (2, 0).
With a simple discrete position update q_new = q + 0.1 times the summed vector, the new point is (1.2, 1). This update is not a validated motor command. The initial potential is 4.5 + 0.25 = 4.75. At the new point it is 3.92 + 0.5625 = 4.4825, so this particular step decreases it.
Now change eta to 6 at the original position. Repulsion becomes (-3, 0), exactly canceling attraction. The same deterministic update remains at (1, 1), although the goal is elsewhere. This demonstrates a stationary point, not proof of a local minimum; in this symmetric example it is a saddle. A route around the point obstacle exists, but the zero local vector does not choose one.

Illustrative case study

In a fictional robotics laboratory, a demonstration works with one repulsion setting and stalls with another. The team initially interprets the stall as a blocked route. Reproducing the cancellation example shows that zero commanded movement can instead arise from the field itself.
They add progress monitoring and test a route-planning fallback that supplies an alternative waypoint. They also replace the point model with geometry representing the vehicle's footprint. Tests cover initial positions away from the symmetry line, narrow passages and control update intervals. The worked calculation explains a local behavior; it does not establish that the revised vehicle can navigate every layout.

Common misconceptions and mistakes

The negative gradient is the downhill direction; using the opposite sign can reverse the intended behavior. A stationary point is not automatically the goal or a local minimum. Local minima are a known problem for ordinary artificial potential fields, but saddles and oscillation can also matter in implementations.
A repulsive term is not, by itself, a universal collision guarantee. Discrete steps, dynamics, geometry and sensing affect the resulting motion. Nor is a potential field identical to A* or hybrid A*: a search method and a local field can be combined, but they solve different parts of a planning design.

Business use

Potential fields can support reactive guidance in a broader robotics system. Specify the robot model, influence distances, gains and meaning of the output vector. Assess completion rate, travel time, clearance and failure recovery across representative environments. Monitor lack of progress explicitly so a stopped controller is not confused with successful arrival. In the example, both an oversized movement step and an overly simplistic obstacle representation would require separate investigation even though the calculated gradient is correct.

Check your understanding

What is the starting summed vector with eta = 2?

Answer: It is (2, 0).

What happens with eta = 6 there?

Answer: The vector becomes zero, so the stated update stalls.

Does decreasing potential on one step prove eventual arrival?

Answer: No; it establishes only that step's decrease.

Related terms

Configuration space
Gradient
Motion planning
Reactive control
Waypoint
obstacle avoidance
local minimum
navigation function
hybrid A*

Learn next

Vary the step size in the worked example and check both potential change and obstacle distance. Then examine why a direction that is geometrically sensible may be infeasible for a vehicle with steering constraints.

Step size

Sources

Lynch and Park, Modern Robotics, virtual potential fields: https://modernrobotics.northwestern.edu/nu-gm-book-resource/10-6-virtual-potential-fields/ Khatib, Real-Time Obstacle Avoidance for Manipulators and Mobile Robots, abstract and publication record: https://khatib.stanford.edu/publications/pdfs/Khatib_1986_IJRR.pdf AIMA glossary, potential field: https://github.com/aimacode/aima-glossary/blob/master/glossary.md

Previous: Postposition Next: Potential scale reduction factor