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.
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.
Instead of a single task, the network learns a family of tasks with many instantiations, which forces generalization rather than memorization.
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.
Lets the network relate the current state to the relevant parts of the demonstration and generalize beyond the training data.
Official
The method generalizes to new instances of task families seen in training; a task outside that family usually requires retraining.
Since one demonstration defines the whole task, its noise, unusual pacing or incomplete execution feed straight through into the policy.
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.
"One-Shot Imitation Learning with Invariance Matching for Robotic Manipulation" demonstrates generalization to new shapes and sim-to-real transfer.
In the Psi-R2.5 model, a single human demonstration converted into robot format suffices for task adaptation without weight updates.
