Shared Modular Recurrence for Universal Morphology Control

ICLR 2026 Conference Submission18962 Authors

Keywords: Deep Reinforcement Learning, Robotic Control, Generalization

TL;DR: Introduction of modular recurrence in the architecture of deep reinforcement learning agents for improved (zero-shot generalization) performance in robotic control.

Abstract: A universal controller for any robot morphology would greatly improve computational and data efficiency. By utilizing contextual information about the properties of individual robots and exploiting their modular structure in the architecture of deep reinforcement learning agents, steps have been made towards multi-robot control. When the robots have highly dissimilar morphologies, this becomes a challenging problem, especially when the agent must generalize to new, unseen robots. In this paper, we hypothesize that the relevant contextual information can be partially observable, but that it can be inferred through interactions for better multi-robot control and generalization to contexts that are not seen during training. To this extent, we implement a modular recurrent transformer-based architecture and evaluate its (generalization) performance on a large set of MuJoCo robots. The results show a substantial improved performance on robots with unseen dynamics, kinematics, and topologies, in four different environments.

Primary Area: reinforcement learning

Submission Number: 18962