Search papers, labs, and topics across Lattice.
This paper introduces the FOCI Policy, an interaction-centric framework that enhances object-centric manipulation policies by focusing on the relative motion between task-relevant objects during short interaction phases. By extracting compact interaction segments from demonstrations and representing skills as relative $SE(3)$ motion, the approach achieves significant generalization while requiring less training data compared to traditional methods. Experiments across multiple platforms, including RLBench and real-world tasks, demonstrate that this method effectively captures interaction dynamics, leading to improved performance in rigid relational manipulation tasks.
Object-object interactions can dramatically reduce the training data needed for effective manipulation policies, achieving superior performance with less complexity.
Object-centric manipulation policies improve generalization by modeling object motion instead of directly predicting robot actions. However, existing methods are often limited by representations which are either too simplistic to capture interaction dynamics or too dense to learn efficiently. We observe that many rigid relational manipulation tasks are governed by short interaction phases where the relative motion between task-relevant objects is tightly constrained. Based on this observation, we propose \textsc{Foci Policy}, an interaction-centric framework that achieves a two-fold abstraction: (1) temporally, by automatically extracting compact interaction segments from demonstrations;(2) spatially, by representing skills as relative $SE(3)$ motion between task-relevant objects, yielding invariance to scene configurations and robot embodiment. Experiments on RLBench, COLOSSEUM, and real-world tasks show that \textsc{Foci Policy} achieves strong performance with substantially less training data than prior object-centric and action-centric policies. These results suggest that modeling object-object interactions provides a simple and efficient inductive bias for rigid relational manipulation. Project page: \href{https://fitz0401.github.io/foci-page/}{fitz0401.github.io/foci-page/}.
