Introduction

In human motion generation and recovery, contact events are frequently used to judge whether interactions with the environment are plausible. This is because meaningful motions are only possible through interactions with the world: from simple everyday motions like walking or picking up an object to more difficult actions such as parkour or gymnastics.

Detecting contact events are usually done via proximity if 3D models are known. Alternatively a model can also be trained to detect contact between body and environment in 2D images or videos. Measuring the magnitude of the resulting forces is a different matter altogether, however. In this post we discuss some common methods of doing so.

Why do we care about the magnitude of the contact forces, especially in the human motion related tasks? While kinematics observations alone are very successful in recreating plausible-looking motions, the lack of dynamics modeling means that rarer motions are difficult to replicate, and generalizations to unseen scenes and interactions suffer. Success in creating plausible trajectories does not imply success in inferring underlying dynamics, as shown in [1]

[1] Vafa, K., Chang, P.G., Rambachan, A., & Mullainathan, S. (2025). What Has a Foundation Model Found? Using Inductive Bias to Probe for World Models.