Human motion capture is often used to guide robot manipulation, but directly mapping human hand movements to a robot hand can break contact with objects. DexOAK, introduced in a new arXiv paper, is a learning-based framework for object-aware kinematic retargeting that generates robot hand references from human motion-capture data.

The method first minimizes an interaction-graph Laplacian, according to the abstract, which appears to encode the relationships between hand parts and objects to maintain contact consistency. The public abstract does not include further details on the network architecture, training data, or experimental results.

The framework aims to produce contact-consistent references for robot hands, a key requirement for reliable manipulation. The paper is available on arXiv under identifier 2610.26762.