Noise-Guided Transport: Imitation Learning from Random Priors

We consider imitation learning in the low-data regime, where only a limited number of expert demonstrations are available. In this setting, methods that rely on large-scale pretraining or high-capacity architectures can be difficult to apply, and efficiency with respect to demonstration data becomes critical. We introduce Noise-Guided Transport (NGT), a lightweight off-policy method that casts imitation as an optimal transport problem solved via adversarial training. NGT requires no pretraining or specialized architectures, incorporates uncertainty estimation by design, and is easy to implement and tune. Despite its simplicity, NGT achieves strong performance on challenging continuous control tasks, including high-dimensional Humanoid tasks, under ultra-low data regimes with as few as 20 transitions.

Learning from demonstration is useful when we want an artificial agent to reproduce skilled behavior, but in many real settings only a few examples are available. This is especially true in areas such as robotics, rehabilitation, and movement analysis, where collecting many high-quality demonstrations can be expensive, slow, or impractical. We introduce Noise-Guided Transport, a method that helps an agent learn from very small demonstration datasets. Instead of directly copying the examples, our method learns a scoring system that tells the agent whether its behavior is close to the demonstrated behavior. It does this by comparing two neural networks: one fixed random network and one trainable network. The difference between them gives a useful signal for separating expert-like behavior from poor behavior. In experiments on challenging simulated control tasks, including humanoid walking, our method learns strong behavior from extremely limited data, sometimes using only a few observed moments of expert behavior. This suggests that imitation learning can remain effective even when demonstrations are scarce, incomplete, or costly to collect. In the long run, methods like this could support learning systems in domains where high-quality examples matter but large datasets are unrealistic.