Autonomous Self-Constructing Modular Robotic Arm using MVMAE-SAC Control for Adaptable Disaster Response

Abstract

Small robotic vehicles can navigate through narrow gaps in rubble but typically lack the strength and dexterity required for manipulation tasks, while larger robotic systems capable of such tasks are often unable to pass through confined spaces that block access to the affected area. The purpose of this project is to investigate whether a low-cost, 3D-printed, modular robotic system can bridge this gap by navigating through confined spaces as separate vehicles before self-assembling into a manipulator, controlled by a reinforcement learning agent operating on only a monocular camera and Time-of-Flight sensor, eliminating the need for expensive sensing hardware such as motor encoders, LiDAR, or depth cameras and reducing overall system cost to a level accessible to underfunded agencies and first responder teams. During training, the camera feed is duplicated and randomly augmented to create multiple views that are encoded using a Multi-View Masked Autoencoder (MVMAE) for spatial representation learning and sim-to-real adaptability. The encoder is trained jointly with a Soft Actor-Critic (SAC) policy, with gradients propagated through the combined encoder reconstruction and policy loss. The agent trained using the MVMAE architecture achieved a 12.4% increase in average episodic reward and a 40.4% reduction in action standard deviation compared to the single-view augmented benchmark, indicating more stable control behavior. The MVMAE-SAC trained model achieved an 89.6% success rate on target-reaching tasks.