AI and Robot Learning
Why Reinforcement Learning Rarely Reaches the Factory
Reinforcement learning dominates robotics research and is almost absent from production lines. Five specific obstacles explain the gap, and only one of them is technical.
Five obstacles, and the decisive ones are not algorithmic. A factory needs predictable behaviour, certifiable safety, diagnosable faults, cheap changes and a clear owner. A learned policy currently offers none of those cleanly, while a taught program offers all five. That is why reinforcement learning appears in locomotion research and almost never on a production line.
The five obstacles
| Requirement | Taught program | Learned policy |
|---|---|---|
| Predictable behaviour | Identical every cycle | Varies with observation noise |
| Certifiable safety | Analysable, provable limits | Empirical competence region only |
| Diagnosable faults | A line of code and a taught point | A weight matrix and a distribution |
| Cheap changes | Reteach a point in minutes | Collect data and retrain |
| Clear ownership | Any controls engineer | A machine learning specialist |
The ownership row is underrated and often decisive. A cell that only one person in the company can modify is a risk that plant managers price correctly, and it is why solutions requiring a specialist to adjust are resisted even when they work.
The reliability gap, in numbers
The comparison a plant makes is not qualitative. It is a set of figures a taught program already meets.
| Metric | Taught program | Learned policy today |
|---|---|---|
| Cycle-to-cycle position variation | ±0.02 to 0.1 mm | 1 to 10 mm |
| Task success in production | 99.9 %+ | 85 to 99 % |
| Time to change a point | 2 to 10 min | hours to weeks |
| People able to modify it | most technicians | 1 to 3 specialists |
| Behaviour after a software update | identical | must be re-evaluated |
| Explanation for a failure | a line number | a distribution shift |
Reading across the first two rows explains the resistance completely. Two orders of magnitude in positional consistency, and one to two orders in failure rate, is not a gap that enthusiasm closes.
Where it does work
- Locomotion. Demonstrations are impractical, simulation is comparatively faithful, and the alternative, hand-tuned gait control, is genuinely worse. This is reinforcement learning's clearest win in robotics.
- Contact-rich insertion research. Learned force strategies outperform hand-written ones for tight-tolerance assembly in laboratory conditions, though production deployment remains rare.
- Process parameter optimisation. Tuning a small number of continuous parameters against a measurable quality outcome, which is a much smaller and safer problem than learning control.
- Scheduling and dispatch. Fleet routing and task assignment, where the action space is discrete and mistakes cost time rather than safety.
Notice what the last two have in common: the learned component sits outside the control loop. That is the pattern that gets past a plant manager, and it is where most production value has actually been realised.
What would change the picture
| Obstacle | What would resolve it | Status |
|---|---|---|
| Predictability | Bounded output guarantees around a nominal trajectory | active research |
| Certification | An accepted method for validating a learned function | standards work under way |
| Diagnosability | Tooling that explains a failure in operator terms | weak |
| Change cost | Fine-tuning from a handful of new demonstrations | improving quickly |
| Ownership | Interfaces a controls engineer can use without retraining | improving |
The fourth row is where progress is fastest. Pretrained policies that adapt from a few hundred demonstrations, rather than requiring a training campaign, directly attack the change-cost objection, which is the one plant engineers raise first.
Practical advice
- Put the learned component outside the control loop wherever possible: perception, selection, scheduling and quality inspection.
- Keep a deterministic fallback for every learned behaviour, and define exactly when it takes over.
- Measure against the conventional solution, not against nothing. The bar is a taught program that already works.
- Choose tasks conventional automation cannot do at all. Mixed depalletising and unstructured picking are the genuine openings; a fixture-located pick is not.
Frequently asked questions
Why is reinforcement learning rare in factories?
Because production needs predictable behaviour, certifiable safety, diagnosable faults, cheap changes and clear ownership. A taught program provides all five and a learned policy currently provides none of them cleanly.
Where does it work in robotics?
Locomotion above all, where demonstrations are impractical and simulation is reasonably faithful. Also process parameter tuning and fleet scheduling, both of which sit outside the control loop.
Can a learned policy be certified?
Not by current accepted methods. Functional safety assessment expects analysable failure modes and provable limits; a trained policy offers empirical evidence instead. Standards work is under way but nothing is settled.
What is the strongest objection from plant engineers?
Change cost, followed by ownership. A cell that requires a machine learning specialist to adjust is a maintenance risk, and a change that requires a retraining campaign rather than a reteach is unacceptable in most production settings.
What should a first deployment look like?
A learned component outside the control loop, a deterministic fallback with a defined handover condition, and a task conventional automation genuinely cannot perform, such as picking from a mixed container.
Sources
- ISO 10218-1:2025, Robotics, Safety requirements, Part 1International Organization for Standardization, functional safety expectations for robot control
- arXiv robotics preprints, reinforcement learning for roboticsPrimary literature on locomotion and contact-rich manipulation results
- World Robotics 2025, industrial robotsInternational Federation of Robotics, the scale of conventional automation this must compete with