MeMo: Meaningful, Modular Controllers via Noise Injection

2024-05-24

Neural Information Processing Systems

AI Summary

To address the inefficiency of training controllers from scratch for novel robotic morphologies, this paper proposes a modular neural controller learning framework. The method decomposes robot control into reusable, semantically meaningful modules via graph-structured representation learning. Its key contributions are: (1) a novel modular objective function that jointly optimizes behavior cloning loss and semantic segmentation of robot subcomponents; (2) a noise injection mechanism that encourages learning of robust, morphology-invariant, component-specific representations; and (3) a graph-based modeling of inter-component relationships to enable both structural and task-level transfer. Experiments on locomotion and dexterous manipulation tasks demonstrate significantly improved training efficiency and superior cross-morphology transfer performance compared to GNN- and Transformer-based baselines. The learned modules exhibit strong semantic interpretability and plug-and-play adaptability across diverse robotic platforms.

Technology Category

• Intelligent Robots: Behavior Learning & Control
• Machine Learning: Multimodal Learning
• Search and Optimization: Learning to Search

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Graph Algorithms and Modeling for the Web: Graph neural networks and deep learning approaches for Web-related graphs
• Responsible Web: Machine-in-the-loop, human agency and autonomy

Abstract

Robots are often built from standardized assemblies, (e.g. arms, legs, or fingers), but each robot must be trained from scratch to control all the actuators of all the parts together. In this paper we demonstrate a new approach that takes a single robot and its controller as input and produces a set of modular controllers for each of these assemblies such that when a new robot is built from the same parts, its control can be quickly learned by reusing the modular controllers. We achieve this with a framework called MeMo which learns (Me)aningful, (Mo)dular controllers. Specifically, we propose a novel modularity objective to learn an appropriate division of labor among the modules. We demonstrate that this objective can be optimized simultaneously with standard behavior cloning loss via noise injection. We benchmark our framework in locomotion and grasping environments on simple to complex robot morphology transfer. We also show that the modules help in task transfer. On both structure and task transfer, MeMo achieves improved training efficiency to graph neural network and Transformer baselines.

Problem

• Develop modular controllers for robot assemblies.
• Enable quick control learning for new robots.
• Improve training efficiency with noise injection.

Innovation

• Modular controllers via noise injection
• Reusable controllers for robot parts
• Improved training efficiency in robotics