Article
Machine learning in robot navigation
Machine learning is used in navigation mainly to interpret sensor data and to anticipate how people move, while the route itself is usually planned by classical algorithms. That split exists because a planner must be predictable and explainable, whereas judging whether a shape ahead is a person is a pattern-recognition problem learning handles well.
Written by Hybot technical lead, Technical lead, Hyrcan-Tech · · 6 min read
Where Hybot stands
Stated first, because it is the part usually left vague: Hybot's platform does not train or run navigation models. Map handling, named points, task assignment and fleet coordination are ours and are deterministic. Perception and motion execution belong to the robot hardware and its manufacturer. Where a given robot uses learned perception, that is the vendor's capability, and we are not going to describe it as if we had built it.
The rest of this article is about the field, not about our product.
Navigation is not one problem
It is at least four, and they have different characters:
| Layer | Question | Typical approach |
|---|---|---|
| Localisation | Where am I on the map? | Probabilistic estimation, sometimes learned components |
| Perception | What is in front of me? | Increasingly learned |
| Global planning | What route should I take? | Classical graph search over a known map |
| Local control | How do I move in the next second? | Classical, with learned prediction sometimes feeding it |
Most "AI navigation" claims refer to the second row. Conflating it with the third is the most common overstatement in the category.
Why perception is where learning earns its place
A sensor return is a set of points. Deciding that a particular cluster of them is a person rather than a pillar, a trolley rather than a wall, is a pattern-recognition problem where hand-written geometry rules become an endless list of exceptions. Learned models are genuinely better at this, and it is not a close contest.
The related win is short-horizon prediction: a person walking across a corridor will probably keep walking. Anticipating that produces smoother behaviour than reacting only when they are already in the way.
Why the planner stays classical
Because when a robot takes an odd route, somebody has to be able to say why.
A deterministic planner over a known map can be interrogated: this edge was blocked, that door was excluded at survey time. A learned policy that produces slightly better routes on average and occasionally an inexplicable one is worse operationally, even if it is better on the benchmark. Predictability is a feature in a room containing customers.
This is the same reasoning that keeps task assignment deterministic in Hybot.
The question to ask a supplier
Not "do you use AI". Ask: which layer is learned, and what does it do when it is uncertain?
A supplier who can answer per layer understands their own system. A supplier who answers "it's AI-powered" has told you nothing you can evaluate.
Where to go next
Cluster hub: AI in robotics.
Frequently asked questions
Which part of navigation benefits most from learning?
- Perception and short-horizon prediction. Deciding whether a sensor return is a person, a pillar or a trolley, and anticipating whether a pedestrian will keep walking, are pattern problems. Classical geometry handles them poorly and learned models handle them well.
Why is route planning still classical?
- Because it must be predictable and inspectable. When a robot takes a strange route, an operator needs an explanation, and a deterministic planner over a known map gives one. A learned policy that is usually better but occasionally inexplicable is a poor trade on a venue floor.
Does Hybot use machine learning for navigation?
- Hybot's platform does not train or run navigation models. Route execution and perception belong to the robot hardware and its manufacturer; our layer handles maps, named points, task assignment and fleet coordination. We will not claim a vendor's perception stack as our own capability.
Where this fits
This page is part of AI in robotics: what is real and what is marketing. If you are working through the topic in order, these are the neighbouring pages.
Computer vision in robotics
What cameras and depth sensors are actually used for on a service robot, what they are bad at, and the privacy questions a venue should ask before installation.
Edge AI in robotics
Why inference on the robot matters when a building's connection is unreliable, what belongs in the cloud instead, and how Hybot's architecture splits the two.
Conversational AI and service robots
The case for and against a talking robot in a noisy venue — what language models are good at, what fails on a restaurant floor, and what Hybot ships instead.
Take it further
If a question here applies to a venue you actually run, the specifics matter more than the general case.