RobotRSI: Toward Full-Stack Recursive Self-Improvement of Robots through Autonomous Research

Abstract

Robots are increasingly able to collect their own experience, update policies, infer models of their bodies, search designs, and participate in automated experiments. In parallel, AI scientists and engineering agents are beginning to formulate hypotheses, write and execute code, coordinate simulation tools, analyse results, and propose revised designs. These developments are usually studied in separate communities and at different layers of the robotics stack. As a result, \emph{self-improving robot} may denote anything from online adaptation of a fixed controller to the automated design and manufacture of a successor machine, while recursive self-improvement is often invoked without evidence that one update made subsequent research more effective. This Survey and Perspective introduces \textbf{RobotRSI}, a unifying framework for autonomous research across robot engineering and for the verified return of research outputs to robots, robot collectives, or successor systems. We organize the field along three orthogonal dimensions: the engineering object being improved, the research activity being automated, and the robot context in which the improvement is produced and evaluated. The engineering scope spans morphology, materials, actuation, sensing, energy, electronics, perception, self- and world models, control, planning, data, learning, simulation, experimentation, manufacturing, deployment, and the research process itself. We distinguish AI scientists for robotics from robot scientists for robotics, and separate the research executor, research target, and improvement recipient. On this basis, we synthesize progress in autonomous hardware, software, and system-level engineering; analyse how constraints vary across general-purpose and specialized robots; and define evidence requirements for local improvement, system update, loop closure, and recursive benefit. The central conclusion is that important local loops already exist, including autonomous policy improvement, self-model revision, morphology search, and physical experimentation, but present evidence is fragmented across system boundaries. Full RobotRSI therefore remains a systems-research agenda rather than a demonstrated monolithic capability. We propose an evaluation framework, reference architecture, and testable roadmap for connecting these local loops without conflating automation, autonomy, embodiment, and recursion.

Keywords