Imitation Learning
Learning a behaviour by copying demonstrations of it, which sidesteps having to specify what success means and inherits whatever the demonstrator did.
Reviewed July 16, 2026Stable
When not to use it
- Where a reward is easy to specify and cheap to evaluate, in which case reinforcement learning explores solutions a demonstrator never tried.
- Where demonstrations are inconsistent, since averaging incompatible strategies produces a policy that follows none of them.
- Where the deployment distribution differs from the demonstration distribution, which is where compounding error does its damage.
Reach for something else instead
- Reinforcement learning — specify a reward and let the system find its own solution, at the cost of needing that reward and far more interaction.
- Inverse reinforcement learning — infer the reward from demonstrations and optimise it, which generalises better and is harder to make work.
- -
Further reading
- Open X-Embodiment Collaboration (2023), Open X-Embodiment: Robotic Learning Datasets and RT-X Models, arXiv:2310.08864 — the pooled corpus and the transfer results.
- Geirhos et al. (2020), Shortcut Learning in Deep Neural Networks — why a policy that matches demonstrations may have learned the demonstration setting.
Primary sources, listed so you can check the claims on this page rather than take them on trust.
Where people go wrong
- Collecting only successful demonstrations, so the policy never learns recovery and small errors cascade.
- Reading demonstration-distribution accuracy as deployment performance, which ignores the distribution the policy itself induces.
- Assuming more demonstrations fix generalisation when the demonstrations are all from one robot in one room.
At a glance
FieldMachine Learning
Simplest formbehavioural cloning
Core failurecompounding error off the demonstration distribution
Alternativeinverse reinforcement learning
Bottleneckdemonstrations, which must be physically produced
DifficultyIntermediate
Where this sits
A starting point. Nothing needs to come before it.
0Levelstarting point
0Needs firstnothing
0Opens upnothing further
1Areastays here
Computed from the prerequisite graph, not assigned. How this works