Few-Shot Imitation: Learning New Tasks from Minimal Demonstrations

Few-shot imitation learning is a paradigm in robot learning where a policy acquires new manipulation skills from as few as 1-10 demonstrations. This capability requires extensive pretraining on a diverse multi-task dataset, during which the model learns general manipulation primitives and the ability to extract task-relevant information from a small number of examples.

The mathematical formulation varies by approach. In meta-learning (MAML), the pretraining objective explicitly optimizes for fast adaptation: the loss function measures performance after a small number of gradient updates on new task data, encouraging the model to learn initializations that are maximally sensitive to new task information. In in-context learning, the demonstrations are simply included as conditioning context in the Transformer's input sequence, and the model learns during pretraining to attend to these demonstrations for task-relevant information.

Few-shot imitation represents a practical sweet spot between zero-shot generalization (which requires no new task demonstrations but achieves lower performance) and full behavioral cloning (which achieves high performance but requires extensive data collection for each new task). For organizations deploying robots across many tasks, few-shot imitation reduces the per-task deployment cost by 10-100x while maintaining reasonable performance levels.

For practitioners, few-shot imitation learning shifts the data collection bottleneck from per-task demonstrations to pretraining dataset diversity. The upfront investment in a diverse multi-task pretraining dataset pays dividends across every subsequent new task. Organizations should focus their data collection budget on maximizing the number of distinct tasks in the pretraining set rather than maximizing demonstrations per task.

The quality of few-shot demonstrations is critical. Each of the 1-10 demonstrations carries outsized influence on the adapted policy's behavior. Demonstrations should be collected by skilled operators, verified for task completion, and reviewed for execution quality. Providing demonstrations with natural variation (slightly different starting positions, different grasp orientations) is better than providing near-identical demonstrations, as variation helps the model identify the task-invariant components.

Claru supports few-shot imitation workflows by providing pre-collected multi-task datasets in standardized formats. These datasets cover diverse manipulation tasks with task-level labeling, enabling direct use as pretraining data. For the few-shot adaptation phase, Claru can also provide small batches of high-quality demonstrations for specific target tasks, collected by trained operators under controlled conditions.

Claru provides the diverse multi-task pretraining datasets that few-shot imitation learning requires, with explicit task labeling across hundreds of manipulation tasks. For the adaptation phase, Claru delivers small batches of high-quality demonstrations collected by skilled operators under controlled conditions.

Few-shot imitation learning is a paradigm where a robot policy learns to perform a new task from a very small number of demonstrations — typically 1 to 10. This contrasts with standard behavioral cloning, which requires hundreds or thousands of demonstrations per task. The 'few-shot' qualifier comes from the broader machine learning concept of few-shot learning, where models generalize from limited examples by leveraging prior knowledge acquired during pretraining.

The technical approach typically involves two phases: a pretraining phase on a large, diverse dataset of many tasks, and an adaptation phase where the pretrained model is conditioned on a few demonstrations of a new target task. During pretraining, the model learns general manipulation primitives, visual representations, and motor skills shared across tasks. During adaptation, the model uses the provided demonstrations as context to specialize its behavior for the specific task without full retraining.

Two main architectural approaches dominate few-shot imitation. Meta-learning methods like MAML (Model-Agnostic Meta-Learning) explicitly optimize the pretraining phase so that a few gradient steps on new task demonstrations produce a high-quality policy. In-context learning methods, used by modern Transformer-based policies, treat the few demonstrations as additional context tokens that condition the policy's predictions without any gradient updates at all. The latter approach is becoming dominant because it requires no optimization at test time — the demonstrations are simply prepended to the observation history.

The quality and diversity of pretraining data is the primary determinant of few-shot imitation performance. A model pretrained on 50 manipulation tasks generalizes better to new tasks in few-shot settings than a model pretrained on 5 tasks, even if the total number of demonstrations is the same. Task diversity during pretraining teaches the model what constitutes a 'task' and how to extract task-relevant features from a small number of demonstrations.

Effective pretraining datasets for few-shot imitation should cover diverse objects (varying geometry, material, weight), diverse actions (reaching, grasping, placing, pushing, pulling, rotating), diverse environments (table heights, background clutter, lighting conditions), and diverse task goals (sorting, stacking, inserting, pouring). Each task should have multiple demonstrations showing natural variation in execution strategy, so the model learns to identify the invariant task structure from variable execution details.

The RLDS format used by Open X-Embodiment provides a natural structure for few-shot pretraining data: each episode is tagged with a task description, enabling the model to group demonstrations by task during training. For few-shot evaluation, held-out tasks with their demonstrations serve as the test set. Claru's multi-task demonstration datasets are collected with explicit task labeling and sufficient per-task variation to support both standard and few-shot training paradigms.

Few-shot imitation is most valuable in settings where tasks change frequently and collecting hundreds of demonstrations per task is impractical. Manufacturing lines that switch between product variants, warehouse operations handling diverse inventory, and assistive robots adapting to individual user preferences all benefit from few-shot capability. The economic argument is straightforward: if a new task requires 5 demonstrations instead of 500, the deployment cost for each new task drops by two orders of magnitude.

The primary limitation of few-shot imitation is that performance on the new task is typically lower than what could be achieved with full-scale data collection. A policy adapted from 5 demonstrations might achieve 70% success rate where 500 demonstrations would yield 95%. For safety-critical tasks (surgical assistance, heavy payload manipulation), this gap may be unacceptable. For lower-stakes tasks (sorting, organizing, simple pick-and-place), the trade-off between data cost and performance is often favorable.

Demonstration quality matters even more in the few-shot setting than in standard behavioral cloning. When the model has only 5 examples to learn from, each one carries significant weight. A single poor demonstration can disproportionately degrade performance. Few-shot demonstrations should be collected by skilled operators under controlled conditions, with careful verification that each demonstration successfully completes the task and represents a clean, efficient execution strategy.