Inductive biases in robot learning for contact detection and collision avoidance

Abstract¶

The terms "model-free" and "model-based" can blur communication between robotics and machine learning communities. For example, "model-free" reinforcement learning trains motion policies in simulation—yet the simulator itself embodies a model. Similarly, in computer vision and natural language processing, models haven't disappeared but evolved into inductive biases: convolutional neural networks and transformers remain foundational architectures in their respective domains. This talk examines how robotics knowledge can be formalized as inductive biases for robot learning, focusing on state estimation and predictive control. We present two case studies—contact detection and obstacle avoidance—showing how combining robotics principles with real-world data creates inductive biases that accelerate training and enable reliable deployment on physical systems.

Content¶

Slides