Learning Dextrous Manipulation Skills for Multifingered Robot Hands Using the Evolution Strategy

We present a method for autonomous learning of dextrous manipulation skills with multifingered robot hands. We use heuristics derived from observations made on human hands to reduce the degrees of freedom of the task and make learning tractable. Our approach consists of learning and storing a few basic manipulation primitives for a few prototypical objects and then using an associative memory to obtain the required parameters for new objects and/or manipulations. The parameter space of the robot is searched using a modified version of the evolution strategy, which is robust to the noise normally present in real-world complex robotic tasks. Given the difficulty of modeling and simulating accurately the interactions of multiple fingers and an object, and to ensure that the learned skills are applicable in the real world, our system does not rely on simulation; all the experimentation is performed by a physical robot, in this case the 16-degree-of-freedom Utah/MIT hand. E xperimental results show that accurate dextrous manipulation skills can be learned by the robot in a short period of time. We also show the application of the learned primitives to perform an assembly task and how the primitives generalize to objects that are different from those used during the learning phase.     

Policy search for motor primitives in robotics

Many motor skills in humanoid robotics can be learned using parametrized motor primitives. While successful applications to date have been achieved with imitation learning, most of the interesting motor learning problems are high-dimensional reinforcement learning problems. These problems are often beyond the reach of current reinforcement learning methods. In this paper, we study parametrized policy search methods and apply these to benchmark problems of motor primitive learning in robotics. We show that many well-known parametrized policy search methods can be derived from a general, common framework. This framework yields both policy gradient methods and expectation-maximization (EM) inspired algorithms. We introduce a novel EM-inspired algorithm for policy learning that is particularly well-suited for dynamical system motor primitives. We compare this algorithm, both in simulation and on a real robot, to several well-known parametrized policy search methods such as episodic REINFORCE, ??Vanilla?? Policy Gradients with optimal baselines, episodic Natural Actor Critic, and episodic Reward-Weighted Regression. We show that the proposed method out-performs them on an empirical benchmark of learning dynamical system motor primitives both in simulation and on a real robot. We apply it in the context of motor learning and show that it can learn a complex Ball-in-a-Cup task on a real Barrett WAM? robot arm.     

Learning how to grasp based on neural network retraining

Humans have an incredible capacity to manipulate objects using dextrous hands. A large number of studies indicate that robot learning by demonstration is a promising strategy to improve robotic manipulation and grasping performance. Concerning this subject we can ask: How does a robot learn how to grasp? This work presents a method that allows a robot to learn new grasps. The method is based on neural network retraining. With this approach we aim to enable a robot to learn new grasps through a supervisor. The proposed method can be applied for 2D and 3D cases. Extensive object databases were generated to evaluate the method performance in both 2D and 3D cases. A total of 8100 abstract shapes were generated for 2D cases and 11700 abstract shapes for 3D cases. Simulation results with a computational supervisor show that a robotic system can learn new grasps and improve its performance through the proposed HRH (Hopfield-RBF-Hopfield) grasp learning approach.     

Systematization of error recovery in skill-based manipulation

Dexterous manipulation is an important function for working robots. Manipulator tasks such as assembly and disassembly can generally be divided into several motion primitives. We call such motion primitives “skills,” and explain how most manipulator tasks can be composed of sequences of these skills. We are currently planning to construct a maintenance robot for household electrical appliances. We considered establishing a hierarchy of the manipulation tasks of this robot since the maintenance of such appliances has become more complex than ever before. In addition, as errors seem likely to increase in complex tasks, it is important to implement an effective error recovery technology. This article presents our proposal for a new type of error recovery that uses the concepts of task stratification and error classification.     

Teacher feedback to scaffold and refine demonstrated motion primitives on a mobile robot

Task demonstration is an effective technique for developing robot motion control policies. As tasks become more complex, however, demonstration can become more difficult. In this work, we introduce an algorithm that uses corrective human feedback to build a policy able to perform a novel task, by combining simpler policies learned from demonstration. While some demonstration-based learning approaches do adapt policies with execution experience, few provide corrections within low-level motion control domains or to enable the linking of multiple of demonstrated policies. Here we introduce Feedback for Policy Scaffolding (FPS) as an algorithm that first evaluates and corrects the execution of motion primitive policies learned from demonstration. The algorithm next corrects and enables the execution of a more complex task constructed from these primitives. Key advantages of building a policy from demonstrated primitives is the potential for primitive policy reuse within multiple complex policies and the faster development of these policies, in addition to the development of complex policies for which full demonstration is difficult. Policy reuse under our algorithm is assisted by human teacher feedback, which also contributes to the improvement of policy performance. Within a simulated robot motion control domain we validate that, using FPS, a policy for a novel task is successfully built from motion primitives learned from demonstration. We show feedback to both aid and enable policy development, improving policy performance in success, speed and efficiency.     

Towards gesture-based programming: shape from motion primordial learning of sensorimotor primitives

Gesture-based programming (GBP) is a paradigm for the evolutionary programming of dextrous robotic systems by human demonstration. We call the paradigm “gesture-based” because we try to capture, in real-time, the intention behind the demonstratrator's fleeting, context-dependent hand motions, contact conditions, finger poses, and even cryptic utterances, rather than just recording and replaying movement. The paradigm depends on a pre-existing knowledge base of capabilities, collectively called “encapsulated expertise”, that comprise the real-time sensorimotor primitives from which the run-time executable is constructed as well as providing the basis for interpreting the teacher's actions during programming. In this paper we first describe the GBP environment, which is not fully implemented. We then present a technique based on principal components analysis, augmentable with model-based information, for learing and recognizing sensorimotor primitives. This paper describes simple applications of the technique to a small mobile robot and a PUMA manipulator. The mobile robot learned to escape from jams while the manipulator learned guarded moves and rotational accommodation that are composable to allow flat plate mating operations. While these initial applications are simple, they demonstrate the ability to extract primitives from demonstration, recognize the learned primitives in subsequent demonstrations, and combine and transform primitives to create different capabilities, which are all critical to the GBP paradigm.     

Object interaction and task programming by demonstration in visuo-haptic augmented reality

A visuo-haptic augmented reality system is presented for object manipulation and task learning from human demonstration. The proposed system consists of a desktop augmented reality setup where users operate a haptic device for object interaction. Users of the haptic device are not co-located with the environment where real objects are present. A three degrees of freedom haptic device, providing force feedback, is adopted for object interaction by pushing, selection, translation and rotation. The system also supports physics-based animation of rigid bodies. Virtual objects are simulated in a physically plausible manner and seem to coexist with real objects in the augmented reality space. Algorithms for calibration, object recognition, registration and haptic rendering have been developed. Automatic model-based object recognition and registration are performed from 3D range data acquired by a moving laser scanner mounted on a robot arm. Several experiments have been performed to evaluate the augmented reality system in both single-user and collaborative tasks. Moreover, the potential of the system for programming robot manipulation tasks by demonstration is investigated. Experiments show that a precedence graph, encoding the sequential structure of the task, can be successfully extracted from multiple user demonstrations and that the learned task can be executed by a robot system.     

Current research in robotics and automation-an intelligent graspingsystem

A research project is described that focuses on building a comprehensive grasping environment capable of performing tasks such as locating moving objects and picking them up, manipulating man-made objects such as tools, and recognizing unknown objects through touch. In addition, an integrated programming environment is being designed that will allow grasping and grasping primitives within an overall robotic control and programming system that includes dextrous hands, vision sensors, and multiple-degree-of-freedom manipulators. A system overview is given, and the applications are discussed     

Team NimbRo at MBZIRC 2017: Autonomous valve stem turning using a wrench

The Mohamed Bin Zayed International Robotics Challenge (MBZIRC) 2017 has defined ambitious new benchmarks to advance the state‐of‐the‐art in autonomous operation of ground‐based and flying robots. In this study, we describe our winning entry to MBZIRC Challenge 2: the mobile manipulation robot Mario. It is capable of autonomously solving a valve manipulation task using a wrench tool detected, grasped, and finally used to turn a valve stem. Mario’s omnidirectional base allows both fast locomotion and precise close approach to the manipulation panel. We describe an efficient detector for medium‐sized objects in three‐dimensional laser scans and apply it to detect the manipulation panel. An object detection architecture based on deep neural networks is used to find and select the correct tool from grayscale images. Parametrized motion primitives are adapted online to percepts of the tool and valve stem to turn the stem. We report in detail on our winning performance at the challenge and discuss lessons learned.     

机器人操作技能学习方法综述

结合人工智能技术和机器人技术,研究具备一定自主决策和学习能力的机器人操作技能学习系统,已逐渐成为机器人研究领域的重要分支.本文介绍了机器人操作技能学习的主要方法及最新的研究成果.依据对训练数据的使用方式将机器人操作技能学习方法分为基于强化学习的方法、基于示教学习的方法和基于小数据学习的方法,并基于此对近些年的研究成果进行了综述和分析,最后列举了机器人操作技能学习的未来发展方向.     

Active learning of inverse models with intrinsically motivated goal exploration in robots

We introduce the Self-Adaptive Goal Generation Robust Intelligent Adaptive Curiosity (SAGG-RIAC) architecture as an intrinsically motivated goal exploration mechanism which allows active learning of inverse models in high-dimensional redundant robots. This allows a robot to efficiently and actively learn distributions of parameterized motor skills/policies that solve a corresponding distribution of parameterized tasks/goals. The architecture makes the robot sample actively novel parameterized tasks in the task space, based on a measure of competence progress, each of which triggers low-level goal-directed learning of the motor policy parameters that allow to solve it. For both learning and generalization, the system leverages regression techniques which allow to infer the motor policy parameters corresponding to a given novel parameterized task, and based on the previously learnt correspondences between policy and task parameters.We present experiments with high-dimensional continuous sensorimotor spaces in three different robotic setups: (1) learning the inverse kinematics in a highly-redundant robotic arm, (2) learning omnidirectional locomotion with motor primitives in a quadruped robot, and (3) an arm learning to control a fishing rod with a flexible wire. We show that (1) exploration in the task space can be a lot faster than exploration in the actuator space for learning inverse models in redundant robots; (2) selecting goals maximizing competence progress creates developmental trajectories driving the robot to progressively focus on tasks of increasing complexity and is statistically significantly more efficient than selecting tasks randomly, as well as more efficient than different standard active motor babbling methods; (3) this architecture allows the robot to actively discover which parts of its task space it can learn to reach and which part it cannot.     

Object–object interaction affordance learning

This paper presents a novel object–object affordance learning approach that enables intelligent robots to learn the interactive functionalities of objects from human demonstrations in everyday environments. Instead of considering a single object, we model the interactive motions between paired objects in a human–object–object way. The innate interaction-affordance knowledge of the paired objects are learned from a labeled training dataset that contains a set of relative motions of the paired objects, human actions, and object labels. The learned knowledge is represented with a Bayesian Network, and the network can be used to improve the recognition reliability of both objects and human actions and to generate proper manipulation motion for a robot if a pair of objects is recognized. This paper also presents an image-based visual servoing approach that uses the learned motion features of the affordance in interaction as the control goals to control a robot to perform manipulation tasks.     

Dextrous hands: human, prosthetic, and robotic

The sensory and motor capacities of the human hand are reviewed in the context of providing a set of performance characteristics against which prosthetic and dextrous robot hands can be evaluated. The sensors involved in processing tactile, thermal, and proprioceptive (force and movement) information are described, together with details on their spatial densities, sensitivity, and resolution. The wealth of data on the human hand's sensory capacities is not matched by an equivalent database on motor performance. Attempts at quantifying manual dexterity have met with formidable technological difficulties due to the conditions under which many highly trained manual skills are performed. Limitations in technology have affected not only the quantifying of human manual performance but also the development of prosthetic and robotic hands. Most prosthetic hands in use at present are simple grasping devices, and imparting a "natural" sense of touch to these hands remains a challenge. Several dextrous robot hands exist as research tools and even though some of these systems can outperform their human counterparts in the motor domain, they are still very limited as sensory processing systems. It is in this latter area that information from studies of human grasping and processing of object information may make the greatest contribution.     

Perceiving,learning, and exploiting object affordances for autonomous pile manipulation

Autonomous manipulation in unstructured environments will enable a large variety of exciting and important applications. Despite its promise, autonomous manipulation remains largely unsolved. Even the most rudimentary manipulation task—such as removing objects from a pile—remains challenging for robots. We identify three major challenges that must be addressed to enable autonomous manipulation: object segmentation, action selection, and motion generation. These challenges become more pronounced when unknown man-made or natural objects are cluttered together in a pile. We present a system capable of manipulating unknown objects in such an environment. Our robot is tasked with clearing a table by removing objects from a pile and placing them into a bin. To that end, we address the three aforementioned challenges. Our robot perceives the environment with an RGB-D sensor, segmenting the pile into object hypotheses using non-parametric surface models. Our system then computes the affordances of each object, and selects the best affordance and its associated action to execute. Finally, our robot instantiates the proper compliant motion primitive to safely execute the desired action. For efficient and reliable action selection, we developed a framework for supervised learning of manipulation expertise. To verify the performance of our system, we conducted dozens of trials and report on several hours of experiments involving more than 1,500 interactions. The results show that our learning-based approach for pile manipulation outperforms a common sense heuristic as well as a random strategy, and is on par with human action selection.     

面向稀疏奖励的机器人操作技能学习

基于深度强化学习的机器人操作技能学习成为研究热点, 但由于任务的稀疏奖励性质, 学习效率较低. 本 文提出了基于元学习的双经验池自适应软更新事后经验回放方法, 并将其应用于稀疏奖励的机器人操作技能学习 问题求解. 首先, 在软更新事后经验回放算法的基础上推导出可以提高算法效率的精简值函数, 并加入温度自适应 调整策略, 动态调整温度参数以适应不同的任务环境; 其次, 结合元学习思想对经验回放进行分割, 训练时动态调整 选取真实采样数据和构建虚拟数的比例, 提出了DAS-HER方法; 然后, 将DAS-HER算法应用到机器人操作技能学 习中, 构建了一个稀疏奖励环境下具有通用性的机器人操作技能学习框架; 最后, 在Mujoco下的Fetch和Hand环境 中, 进行了8项任务的对比实验, 实验结果表明, 无论是在训练效率还是在成功率方面, 本文算法表现均优于其他算 法.     

Automatic grasp planning for multifingered robot hands

In this paper the development of a planning environment is described which was especially tailored for grasping and manipulating with multifinger robot hands. The research has been concerned with the programming and simulation system of the Karlsruhe dextrous hand, which has been in development for two years. The work presents the result of a geometric-mechanic approach to the object-handling problem with dextrous multifinger hands by selecting grasp points and searching grasp forces to perform desired assembly tasks. The knowledge representation for the sequence planning and command execution is based on object and task restrictions combined with routines for successive optimization and a constraint propagation algorithm.     

Task Transfer by a Developmental Robot

Scaffolding is a process of transferring learned skills to new and more complex tasks through arranged experience in open-ended development. In this paper, we propose a developmental learning architecture that enables a robot to transfer skills acquired in early learning settings to later more complex task settings. We show that a basic mechanism that enables this transfer is sequential priming combined with attention, which is also the driving mechanism for classical conditioning, secondary conditioning, and instrumental conditioning in animal learning. A major challenge of this work is that training and testing must be conducted in the same program operational mode through online, real-time interactions between the agent and the trainers. In contrast with former modeling studies, the proposed architecture does not require the programmer to know the tasks to be learned and the environment is uncontrolled. All possible perceptions and actions, including the actual number of classes, are not available until the programming is finished and the robot starts to learn in the real world. Thus, a predesigned task-specific symbolic representation is not suited for such an open-ended developmental process. Experimental results on a robot are reported in which the trainer shaped the behaviors of the agent interactively, continuously, and incrementally through verbal commands and other sensory signals so that the robot learns new and more complex sensorimotor tasks by transferring sensorimotor skills learned in earlier periods of open-ended development     

Demonstration-based learning and control for automatic grasping

We present a method for automatic grasp generation based on object shape primitives in a Programming by Demonstration framework. The system first recognizes the grasp performed by a demonstrator as well as the object it is applied on and then generates a suitable grasping strategy on the robot. We start by presenting how to model and learn grasps and map them to robot hands. We continue by performing dynamic simulation of the grasp execution with a focus on grasping objects whose pose is not perfectly known.