A class of robot learning models can reach the same performance as a standard imitation learning technique using a fraction of the training data, according to new research.
Researchers studied predictive inverse dynamics models (PIDMs) — architectures that pair a future-state predictor with an inverse dynamics model — against behavior cloning (BC), the workhorse method for teaching machines to imitate expert behavior from recorded demonstrations. The paper establishes theoretical conditions under which PIDMs win: the future-state prediction step cuts variance in the model's action estimates, even though it introduces some bias. In 2D navigation tests, BC needed up to five times as many demonstrations as PIDM to hit comparable accuracy, with an average gap of three times. In a 3D video game environment with high-dimensional visual inputs, BC needed more than 66% more samples.
The data efficiency gap matters most when expert demonstrations are expensive to collect — think teleoperated robot arms, surgical simulators, or any domain where labeling is slow and human experts are scarce. The researchers also show the advantage widens when additional unlabeled data sources are available, which is increasingly the common case as robot datasets grow.
Behavior cloning has been the default partly because it is simple and well-understood; this paper does not declare it dead, but it does give practitioners a clearer checklist for when to reach for something more sophisticated instead.