Learning Equivariant Neural-Augmented Object Dynamics From Few Interactions

Problem
This paper addresses the challenge of learning data-efficient object dynamics models for robotic manipulation, particularly for deformable objects. Existing methods, such as modeling objects as sets of 3D particles with graph neural networks (GNNs), often fail to maintain physical feasibility over extended time horizons and require substantial interaction data. The authors propose a novel framework, PIEGraph, to bridge this gap by integrating analytical physics with data-driven approaches, enabling effective learning from limited real-world interactions. This work is presented as a preprint and has not yet undergone peer review.

Method
PIEGraph comprises two main components: (1) a Physically Informed particle-based analytical model, implemented as a spring-mass system, which enforces physically plausible motion, and (2) an Equivariant Graph Neural Network (EGNN) that utilizes a novel action representation to leverage symmetries in particle interactions. The EGNN is designed to enhance the learning of object dynamics by guiding the analytical model with learned representations of object interactions. The training process involves limited interaction data, focusing on both rigid and deformable objects. The architecture’s specifics, including the number of layers and parameters, are not disclosed, but the integration of analytical and learned components is a key innovation.

Results
The authors evaluate PIEGraph on both simulation and real robot hardware, specifically targeting reorientation and repositioning tasks with various objects, including ropes, cloth, stuffed animals, and rigid bodies. PIEGraph demonstrates superior performance compared to state-of-the-art baselines, achieving a significant reduction in prediction error for object dynamics. The results indicate that PIEGraph can maintain physical feasibility while improving the accuracy of dynamics predictions, which is crucial for reliable robotic manipulation planning. Exact numerical results and effect sizes are not provided in the summary, but the authors claim substantial improvements over existing methods.

Limitations
The authors acknowledge several limitations, including the reliance on a specific analytical model (spring-mass system) that may not generalize to all types of deformable objects. Additionally, the performance of PIEGraph may be sensitive to the quality and quantity of the interaction data used for training. The paper does not address potential scalability issues when applied to more complex or diverse object classes, nor does it explore the computational overhead introduced by the integration of analytical and learned components.

Why it matters
The implications of this work are significant for the field of robotic manipulation, particularly in scenarios involving complex object dynamics. By effectively combining analytical physics with data-driven learning, PIEGraph offers a promising approach to enhance the efficiency and reliability of robotic systems in manipulating both rigid and deformable objects. This framework could pave the way for future research on hybrid models that leverage the strengths of both analytical and machine learning techniques, potentially leading to advancements in real-world robotic applications.

Authors: Sergio Orozco, Tushar Kusnur, Brandon May, George Konidaris, Laura Herlant
Source: arXiv:2605.02699
URL: https://arxiv.org/abs/2605.02699v1