In Memory of Alexey V. Borisov. On His 60th Birthday

Implementations of Symmetrical Locomotion in a Quadrupedal Robot with a Neural Processing Unit

A. V. Guba^a, F. R. Khabibullin^a, N. A. Kovalev^a, V. V. Andrulis^a, I. A. Kastalskiy^ba, V. B. Kazantsev<sup>ba

a</sup> Laboratory of Neurobiomorphic Technology, Moscow Institute of Physics and Technology (MIPT), Institutskiy per. 9, Dolgoprudny, Moscow Region, 141701 Russia
^b Department of Neurotechnology, Lobachevsky State University of Nizhny Novgorod (UNN), pros. Gagarina 23, Nizhny Novgorod, 603022 Russia

Abstract: This paper describes a comprehensive approach to the development of a quadrupedal robot possessing 12 actuated degrees of freedom. The development comprises the design of the mechanical system, the creation of control electronics, and the implementation of software for motion generation. A key aspect involves the application of reinforcement learning in a physical simulator, followed by the transfer of the trained algorithms to the physical device (sim-to-real). An embedded Neural Processing Unit (NPU) is utilized to accelerate the execution of AI algorithms, such as object recognition, navigation, and motion optimization. The proposed solutions enable efficient and symmetrical locomotion, high adaptability to changing environmental conditions, and enhanced operational autonomy of the robot.

Keywords: quadruped robot, motion control, physics simulator, reinforcement learning

MSC: 49M25, 68T40

Received: 07.10.2025
Accepted: 09.12.2025

Language: English

DOI: 10.20537/nd251203