Yuya Okadome

Fast Approximation Method for Gaussian Process Regression Using Hash Function for Non-uniformly Distributed Data

Gaussian process regression (GPR) has the ability to deal with non-linear regression readily, although the calculation cost increases with the sample size. In this paper, we propose a fast approximation method for GPR using both locality-sensitive hashing and product of experts models. To investigate the performance of our method, we apply it to regression problems, i.e., artificial data and actual hand motion data. Results indicate that our method can perform accurate calculation and fast approximation of GPR even if the dataset is non-uniformly distributed.

Predictive control method for a redundant robot using a non-parametric predictor

A bio-inspired robot with many degrees of freedom (DOFs) might be beneficial in coping with various situations that occur in a real environment, because its physical structure resembles that of an animal it is modeled after. However, because of its complicated structure, it is difficult to explicitly model the dynamics and to design the control rules. In this study, we propose a predictive control method based on a non-parametric method. Instead of conducting parameter estimation for a certain parametric model, system identification is performed by collecting data. We apply our method to the control of a robot with a complicated structure. Experimental results show that the control of a robot with many DOFs can be achieved by the proposed method. Graphical Abstract

Confidence-based roadmap using Gaussian process regression for a robot control

To achieve a realistic task by a recent complicated robot, a practical motion planning method is important. Especially in this decade, sampling-based motion planning methods have become popular thanks to recent high performance computers. In sampling-based motion planning, a graph that covers the state space is constructed based on reachability between node pairs, and the motion is planned using the graph. However, it requires an explicit model of a controlled target. In this research, we propose a motion planning method in which a system model is estimated by using Gaussian process regression. We apply our method to the control of an actual robot. Experimental results show that the control of the robot can be achieved by the proposed motion planning method.

Estimation of physical interaction between a musculoskeletal robot and its surroundings

Recently, robots are expected to support our daily lives in real environments. In such environments, however, there are a lot of obstacles and the motion of the robot is affected by them. In this research, we develop a musculoskeletal robotic arm and a system identification method for coping with external forces while learning the dynamics of complicated situations, based on Gaussian process regression (GPR). The musculoskeletal robot has the ability to cope with external forces by utilizing a bio-inspired mechanism. GPR is an easy-to-implement method, but can handle complicated prediction tasks. The experimental results show that the behavior of the robot while interacting with its surroundings can be predicted by our method.

HUMA: A human-like musculoskeletal robot platform for physical interaction studies

In a real environment, robots must handle contact with various objects. However, it is hard to model the contacts in advance, since there are a huge variety of objects in our daily lives. Humans have the ability to handle such physical interactions in daily life and such an ability is realized by adapting the physical characteristics produced by the skeletal structure with large DoFs and its actuating system with redundant muscles. In this research, to realize the adaptability of the physical characteristics of a robotic system with commodity type mechanical parts, we developed an actuator network system (ANS) where the motion of multiple actuators are bound by mutually inter-connection. By implementing this system, the response of the robot against various external forces can be modulated or a joint with very large moving range can be realized. In this report, we produced a compliant human-like upper body robot using ANS and investigated the effect of it on the physical characteristic and then the feasibility of the data-driven prediction of the contact force requisite for physical interaction.

Adaptive LSH based on the particle swarm method with the attractor selection model for fast approximation of Gaussian process regression

Gaussian process regression (GPR) is one of the non-parametric methods and has been studied in many fields to construct a prediction model for highly non-linear system. It has been difficult to apply it to a real-time task due to its high computational cost but recent high-performance computers and computationally efficient algorithms make it possible. In our previous work, we derived a fast approximation method for GPR using a locality-sensitive hashing (LSH) and product of experts model, but its performance depends on the parameters of the hash functions used in LSH. Hash functions are usually determined randomly. In this research, we propose an optimization method for the parameters of hash functions by referring to a swarm optimization method. The experimental results show that accurate force estimation of an actual robotic arm is achieved with high computational efficiency.

A confidence-based roadmap using Gaussian process regression

Recent advances in high performance computing have allowed sampling-based motion planning methods to be successfully applied to practical robot control problems. In such methods, a graph representing the local connectivity among states is constructed using a mathematical model of the controlled target. The motion is planned using this graph. However, it is difficult to obtain an appropriate mathematical model in advance when the behavior of the robot is affected by unanticipated factors. Therefore, it is crucial to be able to build a mathematical model from the motion data gathered by monitoring the robot in operation. However, when these data are sparse, uncertainty may be introduced into the model. To deal with this uncertainty, we propose a motion planning method using Gaussian process regression as a mathematical model. Experimental results show that satisfactory robot motion can be achieved using limited data.