Robots Atlas>ROBOTS ATLAS
Training

One-Shot Imitation Learning

2017ActivePublished: 23 September 2026Updated: 23 September 2026Published
Key innovation
Lifting imitation learning to the meta level: the network learns not a single task, but how to read a new task from one demonstration of the same family and execute it immediately.
Category
Training
Abstraction level
Paradigm
Operation level
TrainingRobot control
Use cases
Rapid deployment of new manipulation tasksAssembly and industrial tasks with varying variantsMulti-task robot policiesIn-context learning in robotics foundation modelsGeneralization to new shapes and objects

How it works

Training is done as meta-learning over a family of tasks with many instantiations. The network receives pairs of demonstrations for a subset of tasks: it takes one demonstration plus the current state as input, and its output must reproduce actions matching the trajectory of the second demonstration of the same task. The architecture uses soft attention, which enables generalization beyond the training data and scaling across diverse tasks and conditions. At test time the system is given a single demonstration of a new task and performs on previously unseen instances of it. In modern foundation-model realizations the "input demonstration" takes the form of a robot-format prompt — as in the Psi-R2.5 model, where one human demonstration is converted into robot format and suffices for adaptation without weight updates.

Problem solved

Training a separate policy for every manipulation task is impractical: it demands hundreds of demonstrations per task or manual feature engineering. One-shot imitation learning covers an entire task family with one network and adds a new task at the cost of a single demonstration.

Components

Meta-learning over task familiesTraining regime

Instead of a single task, the network learns a family of tasks with many instantiations, which forces generalization rather than memorization.

Demonstration pairs as the learning signalTraining objective construction

The input is one demonstration plus the current state; the target is actions consistent with the trajectory of a second demonstration of the same task.

Soft attentionGeneralization mechanism

Lets the network relate the current state to the relevant parts of the demonstration and generalize beyond the training data.

Official

Implementation

Reference implementations
IMOP — One-Shot Imitation Learning with Invariance Matching
Python · Authors of the RSS 2024 paper
Implementation pitfalls
Generalization only within the familyHigh

The method generalizes to new instances of task families seen in training; a task outside that family usually requires retraining.

Fix:Design the training task family to cover the intended deployment space.
Sensitivity to the quality of the single demonstrationMedium

Since one demonstration defines the whole task, its noise, unusual pacing or incomplete execution feed straight through into the policy.

Fix:Validate the input demonstration and, where possible, convert it into robot format through a validated pair model.

Evolution

Original paper · 2017 · arXiv preprint (2017) · Yan Duan
One-Shot Imitation Learning
Yan Duan, Marcin Andrychowicz, Bradly C. Stadie, Jonathan Ho, Jonas Schneider, Ilya Sutskever, Pieter Abbeel, Wojciech Zaremba
2017
Publication of "One-Shot Imitation Learning"
Inflection point

Duan and co-authors formulate the paradigm: meta-learning over task families, input consisting of one demonstration plus state, and soft attention as the generalization mechanism.

2024
Invariance matching (IMOP) at RSS 2024

"One-Shot Imitation Learning with Invariance Matching for Robotic Manipulation" demonstrates generalization to new shapes and sim-to-real transfer.

2026
Convergence with in-context learning in foundation models

In the Psi-R2.5 model, a single human demonstration converted into robot format suffices for task adaptation without weight updates.