The results of structural-parametric synthesis of planar underactuated linkages with variable length links are presented. The proposed synthesis is based on the principle of morphological computation, which means that robot’s desired properties are „programmed“ at mechanical level, while active control is used only for soft motion correction caused by the natural behavior of the system. The proposed synthesis algorithm was tested to create an adaptive compact hand for the anthropomorphic robot iCub, capable of performing all fundamental grips and having a higher load capacity with dimensions similar to the original cable-driven hand design. An example of the synthesis for an index finger mechanism is given and the results of testing a prototype device are presented. The developed prototype has 14 degrees of freedom, but it is controlled by just four motors. Minimizing the number of motors made it possible to simplify the hardware and software of its control system, reduce the number of required sensors and weight and size parameters and reduce the cost of components.

The results of the development and implementation of hardware and software architecture for the control system of an adaptive anthropomorphic robotic hand are presented. The research focuses on possibility of integrating the proposed robotic hand into the iCub robot’s control system while preserving the functionality and flexibility of implementing control algorithms. It has been achieved by prototyping the control system as an independent module connected to the iCub robot via network interface Ethernet. The data exchange between the gripper and the iCub robot has high stability and performance with a control frequency of 2 kHz with a delay less than 310 us and a jitter below 50 us. Testing of the control system’s software and hardware architecture demonstrated high accuracy in position control (± 1⁰) and force control (± 0,15 N) for fingers’ proximal phalanges.

Results of a study on the design and manufacture of a prototype of an energy-efficient jumping robot with flexible joints using the principles of morphological calculation are presented. Flexible elements allow robots adaptation to the environment during contact interaction, redirecting the interaction energy from the plastic deformation of solids to the elastic deformation of elastic bodies, which contributes to energy recovery in the system. Unlike traditional lower and higher kinematic pairs, flexible joints provide movement of links only in a limited range within the elastic deformation zone. The problem of designing elastic polymer cross joints is solved by the example of a flat leg mechanism of an incomplete jumping robot of closed kinematics, driven by a single servo motor with elastic elements connected in series. When synthesizing such a robot, it is necessary to optimize not only the kinematic parameters of the lever mechanism, but also the topology and elastic-static parameters of the elastic joints themselves.

An algorithm for planning the movement of a multi-link robotic system in an environment with obstacles is considered. The main requirements for this task are high performance and efficient memory usage during operation. An algorithm for path planning based on the method of a bidirectional fast-investigating random tree is presented. An approach is used which excludes addition of new vertices to the tree if their location in space can unambiguously conclude that it is inappropriate to use them in the path construction. This modification makes it possible to speed up movement planning and reduce the amount of memory needed to store the environment analysis data.

Ways to expand the set of recognized object classes for the task of segmenting them, where it is necessary to build an object mask, as well as to find out its class, are considered. For the first task, methods that do not depend on the classes of subjects and are the most resistant to shape changes were used; for the second task, methods based on iterative learning and methods of metric learning are analyzed. The second approach is chosen as the main one, and various neural network architectures are tested for it. The classification of objects using the k nearest neighbors algorithm is carried out. The COIL-100 set is used as a data set for training a neural network, and after that the trained model was tested on its own data set. The experiments show that the method used allows processing 7-8 images per second on a GTX 1050 ti graphics card with 4 GB of video memory with a classification accuracy of 99%.

Results of a study aimed at developing an intelligent autonomous navigation system for warehouse and office logistics using deep neural networks, are presented. The modern and most versatile tools for depth maps retrieval and semantic data segmentation on images in different environments are analyzed. A comparison of depth maps retrieved hardware from RGB-D camera, neural network algorithms, and a modified Hirschmuller algorithm is carried out. Results of testing performed with a specially prepared dataset shot in an office space, including many complex objects such as glass, mirrors, and multiple light sources demonstrate that the proposed solution outperforms the alternatives in accuracy and uses fewer computational resources in the process.

The problem of improving the accuracy and robustness of simultaneous localization and mapping methods using numerical optimization with constraints is considered. The proposed solution is based on a modification of the ORB-SLAM3 algorithm, which takes into account the peculiarities of the kinematics of ground robots and uses the complexing of visual and wheel odometry data, bundle adjustment for setting parameters that comprehensively characterize the state of the visual sensor, as well as the loop closure algorithm to correct the map. Results of the approach approbation with an OpenLoris dataset demonstrate that for several scenarios the proposed solution is significantly superior in accuracy and robustness to the ORB-SLAM3 algorithm.