Search

Robot learning by demonstration is a research paradigm that can play an important role in expanding areas where robots can be applied. An easy method to get a policy to reproduce the demonstrated behavior is to learn a model that maps directly from demonstrations to action. This is referred to as behavioral cloning (BC). In this paper, a novel robotic BC method for directly imitating a policy from trajectories generated by performing a task is proposed. The proposed method not only allows a user to obtain a demonstrated trajectory for performing a task by manipulating the robot, but also automatically generates a trajectory by having the robot perform the task on behalf of the user. Experiment results showed that the proposed method can effectively learn to make accurate behavior prediction for robot manipulators.

Recently, reinforcement learning utilizing deep learning, called deep reinforcement learning, is being developed for discontinuous action space. It is shown that deep reinforcement learning has higher playing skills than those of human in the video game as Atari. The development of deep reinforcement learning for discontinuous action space, has been extended to the continuous action space as robot control area. The area of control deals with issues closely related to real world environment. However, deep reinforcement learning has limitations to solve control problems in the real world due to numerous trial-and-error based training. To address the limitations of learning in the real world, virtual world environments are used to emulate control problem in real world. In this paper, to solve a complex robot manipulation task, the learning robot manipulation based modular reward shaping is proposed. Since it is difficult to solve the complex task using only single primitive manipulation skill with a reward, it is required to combine various primitive manipulation skills with various rewards. In simulation results, it is shown that the combination of the reward functions affects to the robot manipulation performance.

Advances in deep reinforcement learning enabled cost-efficient training of control policy of physical robot actions from robot simulators. Learning control policy in a simulated environment is cost-efficient over learning in a real environment. Reward engineering is one of the key components to train efficient control policy. For tasks with long horizons such as navigation and manipulation, a sparse reward is providing limited information. The robot simulator for a physical engine of physical robot manipulation has made it easy for researchers in the field of deep reinforcement learning to simulate complicated robot manipulation environments. In this paper, A robot manipulation simulator and a deep RL framework are utilized for implement a training control policy by utilizing previous experiences. For implementation, Recent innovation Hindsight Experience Replay (HER) algorithms with previous experiences to calculate dense rewards from a sparse reward is leveraged . Proposed implementation showed an approach to investigate the reward engineering method to formulate dense reward in robot manipulator tasks.

This paper provides an experimental study of reinforcement learning on IoT devices using distilled knowledge, whose a teacher with a well-trained model transfers to a student with a new model to be trained. The experimental results show that the distilled knowledge is effective to a new model training on IoT devices.

Developing an autonomous indoor mobile robot navigation system from scratch is very difficult and it takes a long time. It requires a series of complex tasks such as handling given sensor inputs, calculating all the information needed for autonomous navigation, and controlling actuators required for movement. In this paper, an autonomous navigation system for indoor mobile robots is introduced mainly based on open source provided by the robot operating system. The presented system is capable of autonomously navigating an unstructured indoor environment avoiding collision with static or dynamic objects. To this end, the system consists of three main modules: mapping, localization, and planning. The mapping module builds a global map for an unknown environment by means of a simultaneous localization and mapping algorithm based on laser scanner data. The localization module estimates the mobile robot’s pose within the prebuilt map by way of an adaptive Monte Carlo localization approach. The planning module builds a local cost map for collision avoidance, generates collision-free trajectories to reach a goal pose based on the cost map, and produces control commands to follow the trajectories. The presented system has been tested not only in simulation environments built in the Gazebo simulator but also in real environments utilizing the Jackal mobile robot, to validate its performance for autonomous navigation including collision avoidance.

Today, a paradigm in the manufacturing industry has been shifted from mass production to small and custom production. To adapt to today’s industrial expectations, this paper presents position-based visual servoing of multiple robotic manipulators. The presented visual servoing consists of two modules: task planning and motion planning. The task planning module manipulates each robotic arm not to collide with each other by distributing a task to them in a systematic manner. The motion planning module employs the MoveIt framework to solve inverse kinematics for each robotic arm so that it moves each end-effector’s pose to the desired pose to complete the task. In addition, the module controls a robotic hand to grasp or release a target object. The presented visual servoing has been tested and verified in the Gazebo simulator.

Applying deep learning to symbolic planning is not straightforward because they differently internalize knowledge. One represents knowledge numerically while the other does symbolically. Hybrid methods that take strengths from both techniques will deliver better performance on automated planning. This paper presents a method to improve the performance of symbolic planning by applying LSTM recurrent networks to inferring the guidance knowledge for search tree node expansion. The networks learn sequential patterns across previous node expansions and generalizes the learned knowledge to predict promising actions and their probability scores at the present expansion step. The prediction consists with node cost estimation for heuristic search. A simple experiment vindicates the method of this research.

This paper addresses a continual lifelong learning problem that learns incremental multiple tasks in real-world environments. We overview and summarize representative approaches and categorization of the state-of-the-art in continual learning. Comparable scenarios, benchmark datasets, and baseline approaches for different continual scenarios introduced in this paper. We suggested a comparison of the differences and similarities with other machine learning methods. We also report real-world applications, especially robots and healthcare fields. We summarize current states and suggest future direction of continual learning problems.