Beyond Knowing — Deciding

A robot that knows where it is and what is around it still has to decide what to do. Should it turn left or right around that obstacle? Slow down for the child running across the room? Skip the messy area entirely and come back later? Motion planning is the umbrella term for all the algorithms that answer these questions.

The Classical Approach

For decades, motion planning meant graph search algorithms — primarily Dijkstra’s algorithm and A* (pronounced “A-star”). The robot models its environment as a grid of cells, marks obstacles as impassable, and searches for the cheapest path from start to goal. This works beautifully for static environments.

Robotic lawn mowers like the Mammotion Luba 2 AWD 5000 still rely heavily on classical planning because lawns are mostly static — the grass does not move, and the boundaries do not change between mowing sessions.

The Modern Problem: Dynamic Environments

Homes, sidewalks, warehouses — almost all real-world environments are dynamic. People move, pets wander, doors open and close, furniture gets rearranged. Classical algorithms struggle here because they assume the world stays put while you plan.

Modern motion planners run the entire planning pipeline 10-50 times per second, continuously replanning as new information arrives from the robot’s vision system and SLAM.

Three Approaches in Modern Use

Sampling-Based Planners

Instead of evaluating every possible move, sampling-based planners (RRT, RRT*, PRM) randomly throw out candidate movements and check which ones are valid and lead toward the goal. They scale much better than grid-based methods for robots with many degrees of freedom — exactly what is needed for humanoid robots like Figure 02 or Ameca.

Optimisation-Based Planners

Model Predictive Control (MPC) treats motion as an optimisation problem: find the sequence of movements that minimises some cost function (energy, time, distance to obstacles) over the next 1-2 seconds. Then execute the first movement, re-plan, and repeat. This is what most self-driving cars use.

Learned Policies

The newest approach is reinforcement learning. Robots learn motion policies by trial and error in simulation, then deploy the learned policy directly. This is how Boston Dynamics’ robots achieve such fluid, almost athletic motion — they learned by literally falling over a few hundred thousand times in simulation first.

The Hidden Heuristic: Cost Functions

The “intelligence” of a motion planner mostly lives in its cost function. Planners do not just optimise for shortest distance — they balance many competing concerns:

  • Distance from obstacles (further is safer but longer)
  • Travel time (fastest path may not be safest)
  • Energy use (smoother motions save battery)
  • Comfort (in passenger applications, jerk and acceleration matter)
  • Social conventions (delivery robots learn to pass pedestrians on the correct side)

What Goes Wrong

Motion planning failures are often surprising. A robot might freeze in a doorway because two acceptable paths cancel each other out in the cost function. A delivery robot might pace back and forth waiting for a “good enough” gap in pedestrians that never comes. These edge cases are what separate research demos from products people actually trust.

The Direction Things Are Heading

The next frontier is foundation models for motion. Just as language models learn general-purpose language from huge datasets, researchers are training general-purpose motion models on massive datasets of robot behaviour. Early results suggest these models can transfer surprisingly well between robot bodies — a policy learned on one humanoid can work on another with minimal adjustment.

That capability, if it scales, could compress what is currently a per-product engineering effort into a software download.