Embedding Physics Priors in Robot Learning: A Survey

Embedding Physics Priors in Robot Learning. A Survey. By Mattia Piccinini and twelve others, from the Technical University of Munich, the Technical University of Darmstadt, and two other institutions. Published on the fifteenth of September. Its code repository is public but slight: six source files.

Artificial intelligence is moving fast into robotics, and learning-based approaches are being adopted quickly. Purely data-driven methods have done remarkably well in computer vision and natural language processing. Robotics is harder going. Data are limited, real-world interactions are complex, and systems have to operate reliably. Those pressures have pushed researchers toward physics-embedded robot learning, which builds physics priors into learning algorithms. The idea is to stop asking a learning system to rediscover, from scarce examples, laws that are already known.

The appeal is that physical laws and constraints, once encoded, can make up for limited data with inductive biases specific to robotics. That could improve generalization, interpretability and sample efficiency. But the literature is fragmented across terminology, methods and application areas, which makes the growing body of work hard to assess as a whole. Similar ideas travel under different names, and work in one application area rarely speaks to work in another. A survey that imposes one vocabulary is, in that setting, a contribution in its own right.

This survey covers a broad span: many kinds of physics prior, many robotics applications, and machine learning models from single-layer perceptrons all the way to generative foundation models. Its organising device is a unified taxonomy that classifies approaches by how physics is embedded. There are three routes. Physics can guide the inputs, the data and the representations. It can be encoded into the model architecture. Or it can inform the training loss function.

On top of that taxonomy, the survey reviews methods for robot dynamics learning, trajectory planning, prediction, control and estimation. It also covers the open-source software ecosystem that supports this work. Crossing the taxonomy with these task areas gives a grid: for any robotics problem, one can ask which of the three kinds of embedding has been tried on it.

The authors identify key open challenges and outline promising directions for future research. Their overall position is stated plainly. Physics priors are a particularly relevant inductive bias for robotics, and they complement data-driven learning rather than replacing it. That framing matters. The survey is not arguing against learning from data. It is arguing that, in a field where data are scarce and reliability is required, known physics is too useful to leave out.

The three-way taxonomy is the part a reader can carry away and reuse. Asking of any method whether its physics lives in the data, in the architecture, or in the loss makes very different approaches comparable. The authors see this path leading toward robotic systems that are more generalizable, more data-efficient, and more trustworthy.