Inductive biases in robot learning for contact detection and collision avoidance

Abstract¶

The overlapping terminology between the robotics and machine learning communities sometimes creates misunderstandings. "Model‑free" reinforcement learning, for instance, is widely used to train motion policies in simulation and appears to contradict classic "model‑based" approaches. In reality, the simulator itself provides an implicit model. Likewise, in other fields revisited by machine learning, rather than vanishing models have been displaced as "inductive biases": convolutional neural networks continue to play a pivotal role in computer vision, as well as transformers in natural language processing. In this talk, we ask: what knowledge from robotics science can we formalize as inductive biases for robot learning? We explore this question through the lenses of state estimation and predictive control, focusing on the two problems of contact detection and obstacle avoidance. By combining insights from robotics with real-world data, we will discuss inductive biases that accelerate training while allowing deployment on real systems.

Content¶

Slides