Learning person-specific models for facial expression and action unit recognition

A key assumption of traditional machine learning approach is that the test data are draw from the same distribution as the training data. However, this assumption does not hold in many real-world scenarios. For example, in facial expression recognition, the appearance of an expression may vary significantly for different people. As a result, previous work has shown that learning from adequate person-specific data can improve the expression recognition performance over the one from generic data. However, person-specific data is typically very sparse in real-world applications due to the difficulties of data collection and labeling, and learning from sparse data may suffer from serious over-fitting. In this paper, we propose to learn a person-specific model through transfer learning. By transferring the informative knowledge from other people, it allows us to learn an accurate model for a new subject with only a small amount of person-specific data. We conduct extensive experiments to compare different person-specific models for facial expression and action unit (AU) recognition, and show that transfer learning significantly improves the recognition performance with a small amount of training data.     

Supervised fuzzy reinforcement learning for robot navigation

This paper addresses a new method for combination of supervised learning and reinforcement learning (RL). Applying supervised learning in robot navigation encounters serious challenges such as inconsistent and noisy data, difficulty for gathering training data, and high error in training data. RL capabilities such as training only by one evaluation scalar signal, and high degree of exploration have encouraged researchers to use RL in robot navigation problem. However, RL algorithms are time consuming as well as suffer from high failure rate in the training phase. Here, we propose Supervised Fuzzy Sarsa Learning (SFSL) as a novel idea for utilizing advantages of both supervised and reinforcement learning algorithms. A zero order Takagi–Sugeno fuzzy controller with some candidate actions for each rule is considered as the main module of robot's controller. The aim of training is to find the best action for each fuzzy rule. In the first step, a human supervisor drives an E-puck robot within the environment and the training data are gathered. In the second step as a hard tuning, the training data are used for initializing the value (worth) of each candidate action in the fuzzy rules. Afterwards, the fuzzy Sarsa learning module, as a critic-only based fuzzy reinforcement learner, fine tunes the parameters of conclusion parts of the fuzzy controller online. The proposed algorithm is used for driving E-puck robot in the environment with obstacles. The experiment results show that the proposed approach decreases the learning time and the number of failures; also it improves the quality of the robot's motion in the testing environments.     

Learning complex action models with quantifiers and logical implications

Automated planning requires action models described using languages such as the Planning Domain Definition Language (PDDL) as input, but building action models from scratch is a very difficult and time-consuming task, even for experts. This is because it is difficult to formally describe all conditions and changes, reflected in the preconditions and effects of action models. In the past, there have been algorithms that can automatically learn simple action models from plan traces. However, there are many cases in the real world where we need more complicated expressions based on universal and existential quantifiers, as well as logical implications in action models to precisely describe the underlying mechanisms of the actions. Such complex action models cannot be learned using many previous algorithms. In this article, we present a novel algorithm called LAMP (Learning Action Models from Plan traces), to learn action models with quantifiers and logical implications from a set of observed plan traces with only partially observed intermediate state information. The LAMP algorithm generates candidate formulas that are passed to a Markov Logic Network (MLN) for selecting the most likely subsets of candidate formulas. The selected subset of formulas is then transformed into learned action models, which can then be tweaked by domain experts to arrive at the final models. We evaluate our approach in four planning domains to demonstrate that LAMP is effective in learning complex action models. We also analyze the human effort saved by using LAMP in helping to create action models through a user study. Finally, we apply LAMP to a real-world application domain for software requirement engineering to help the engineers acquire software requirements and show that LAMP can indeed help experts a great deal in real-world knowledge-engineering applications.     

Learning object deformation models for robot motion planning

In this paper, we address the problem of robot navigation in environments with deformable objects. The aim is to include the costs of object deformations when planning the robot’s motions and trade them off against the travel costs. We present our recently developed robotic system that is able to acquire deformation models of real objects. The robot determines the elasticity parameters by physical interaction with the object and by establishing a relation between the applied forces and the resulting surface deformations. The learned deformation models can then be used to perform physically realistic finite element simulations. This allows the planner to evaluate robot trajectories and to predict the costs of object deformations. Since finite element simulations are time-consuming, we furthermore present an approach to approximate object-specific deformation cost functions by means of Gaussian process regression. We present two real-world applications of our motion planner for a wheeled robot and a manipulation robot. As we demonstrate in real-world experiments, our system is able to estimate appropriate deformation parameters of real objects that can be used to predict future deformations. We show that our deformation cost approximation improves the efficiency of the planner by several orders of magnitude.     

Learning to select distinctive landmarks for mobile robot navigation

In landmark-based navigation systems for mobile robots, sensory perceptions (e.g., laser or sonar scans) are used to identify the robot’s current location or to construct internal representations, maps, of the robot’s environment. Being based on an external frame of reference (which is not subject to incorrigible drift errors such as those occurring in odometry-based systems), landmark-based robot navigation systems are now widely used in mobile robot applications.The problem that has attracted most attention to date in landmark-based navigation research is the question of how to deal with perceptual aliasing, i.e., perceptual ambiguities. In contrast, what constitutes a good landmark, or how to select landmarks for mapping, is still an open research topic. The usual method of landmark selection is to map perceptions at regular intervals, which has the drawback of being inefficient and possibly missing ‘good’ landmarks that lie between sampling points.In this paper, we present an automatic landmark selection algorithm that allows a mobile robot to select conspicuous landmarks from a continuous stream of sensory perceptions, without any pre-installed knowledge or human intervention during the selection process. This algorithm can be used to make mapping mechanisms more efficient and reliable. Experimental results obtained with two different mobile robots in a range of environments are presented and analysed.     

Using roadmaps to classify regions of space for autonomous robot navigation

The roadmap approach to robot path planning is one of the earliest methods. Since then, many different algorithms for building roadmaps have been proposed and widely implemented in mobile robots but their use has always been limited to planning in static, totally known environments. In this paper we combine the use of dynamic analogical representations of the environment with an efficient roadmap extraction method, to guide the robot navigation and to classify the different regions of space in which the robot moves. The paper presents the general reference architecture for the robotic system and then focuses on the algorithms for the construction of the roadmap, the classification of the regions of space and their use in robot navigation. Experimental results indicate the applicability and robustness of this approach in real situations.     

A virtual environment-based system for the navigation of underwater robots

Efficient teleoperation of underwater robot requires clear 3D visual information of the robot's spatial location and its surrounding environment. However, the performance of existing telepresence systems is far from satisfactory. In this paper, we present our virtual telepresence system for assisting tele-operation of an underwater robot. This virtual environment-based telepresence system transforms robot sensor data into 3D synthetic visual information of the workplace based on its geometrical model. It provides the operators with a full perception of the robot's spatial location. In addition, we propose a robot safety domain to overcome the robot's location offset in the virtual environment caused by its sensor errors. The software design of the system and how a safety domain can be used to overcome robot location offset in virtual environment will be examined. Experimental tests and its result analysis will also be presented in this paper.     

Reinforcement based mobile robot navigation in dynamic environment

In this paper, a new approach is developed for solving the problem of mobile robot path planning in an unknown dynamic environment based on Q-learning. Q-learning algorithms have been used widely for solving real world problems, especially in robotics since it has been proved to give reliable and efficient solutions due to its simple and well developed theory. However, most of the researchers who tried to use Q-learning for solving the mobile robot navigation problem dealt with static environments; they avoided using it for dynamic environments because it is a more complex problem that has infinite number of states. This great number of states makes the training for the intelligent agent very difficult. In this paper, the Q-learning algorithm was applied for solving the mobile robot navigation in dynamic environment problem by limiting the number of states based on a new definition for the states space. This has the effect of reducing the size of the Q-table and hence, increasing the speed of the navigation algorithm. The conducted experimental simulation scenarios indicate the strength of the new proposed approach for mobile robot navigation in dynamic environment. The results show that the new approach has a high Hit rate and that the robot succeeded to reach its target in a collision free path in most cases which is the most desirable feature in any navigation algorithm.     

Discrete nonlinear observers for inertial navigation

We derive an exact deterministic nonlinear observer to compute the continuous state of an inertial navigation system based on partial discrete measurements, the so-called strap-down problem. Nonlinear contraction is used as the main analysis tool, and the hierarchical structure of the system physics is systematically exploited. The paper also discusses the use of nonlinear measurements, such as distances to time-varying reference points.     

Live semantic data from building digital twins for robot navigation: Overview of data transfer methods

Increasing reliance on automation and robotization presents great opportunities to improve the management of construction sites as well as existing buildings. Crucial in the use of robots in a built environment is their capacity to locate themselves and navigate as autonomously as possible. Robots often rely on planar and 3D laser scanners for that purpose, and building information models (BIM) are seldom used, for a number of reasons, namely their unreliability, unavailability, and mismatch with localization algorithms used in robots. However, while BIM models are becoming increasingly reliable and more commonly available in more standard data formats (JSON, XML, RDF), they become more promising and reliable resources for localization and indoor navigation, in particular in the more static types of existing infrastructure (existing buildings). In this article, we specifically investigate to what extent and how such building data can be used for such robot navigation. Data flows are built from BIM model to local repository and further to the robot, making use of graph data models (RDF) and JSON data formats. The local repository can hereby be considered to be a digital twin of the real-world building. Navigation on the basis of a BIM model is tested in a real world environment (university building) using a standard robot navigation technology stack. We conclude that it is possible to rely on BIM data and we outline different data flows from BIM model to digital twin and to robot. Future work can focus on (1) making building data models more reliable and standard (modelling guidelines and robot world model), (2) improving the ways in which building features in the digital building model can be recognized in 3D point clouds observed by the robots, and (3) investigating possibilities to update the BIM model based on robot feedback.     

Neural networks for the EMOBOT robot control architecture

Within the EMOBOT approach to adaptive behaviour, the task of learning to control the behaviour is one of the most interesting challenges. Learned action selection between classically implemented control mechanisms, with respect to internal values and sensor readings, provides a way to modulate a variety of behavioural capabilities. To demonstrate the potential of the learning emotional controller, we chose a 10-5-12 MLP to implement the , controller of the EMOBOT. Since no teacher vector is available for the chosen task, the neural network is trained with a reinforcement strategy. The emotion-value-dependent reinforcement signal, together with the output of the network, is the basis with which to compute an artificial teacher vector. Then, the established gradient descent method (backpropagation of error) is applied to train the neural network. First results obtained by extensive simulations show that a still unrevealed richness in behaviour can be realised when using the neural-network-based learning emotional controller.     

Incremental Learning of Linear Model Trees

A linear model tree is a decision tree with a linear functional model in each leaf. Previous model tree induction algorithms have been batch techniques that operate on the entire training set. However there are many situations when an incremental learner is advantageous. In this article a new batch model tree learner is described with two alternative splitting rules and a stopping rule. An incremental algorithm is then developed that has many similarities with the batch version but is able to process examples one at a time. An online pruning rule is also developed. The incremental training time for an example is shown to only depend on the height of the tree induced so far, and not on the number of previous examples. The algorithms are evaluated empirically on a number of standard datasets, a simple test function and three dynamic domains ranging from a simple pendulum to a complex 13 dimensional flight simulator. The new batch algorithm is compared with the most recent batch model tree algorithms and is seen to perform favourably overall. The new incremental model tree learner compares well with an alternative online function approximator. In addition it can sometimes perform almost as well as the batch model tree algorithms, highlighting the effectiveness of the incremental implementation. Editor: Johannes Fürnkranz     

Adaptive learning codebook for action recognition

Learning a compact and yet discriminative codebook is an important procedure for local feature-based action recognition. A common procedure involves two independent phases: reducing the dimensionality of local features and then performing clustering. Since the two phases are disconnected, dimensionality reduction does not necessarily capture the dimensions that are greatly helpful for codebook creation. What’s more, some dimensionality reduction techniques such as the principal component analysis do not take class separability into account and thus may not help build an effective codebook. In this paper, we propose the weighted adaptive metric learning (WAML) which integrates the two independent phases into a unified optimization framework. This framework enables to select indispensable and crucial dimensions for building a discriminative codebook. The dimensionality reduction phase in the WAML is optimized for class separability and adaptively adjusts the distance metric to improve the separability of data. In addition, the video word weighting is smoothly incorporated into the WAML to accurately generate video words. Experimental results demonstrate that our approach builds a highly discriminative codebook and achieves comparable results to other state-of-the-art approaches.     

Neuro-fuzzy techniques to optimize an FPGA embedded controller for robot navigation

This paper describes how low-cost embedded controllers for robot navigation can be obtained by using a small number of if-then rules (exploiting the connection in cascade of rule bases) that apply Takagi–Sugeno fuzzy inference method and employ fuzzy sets represented by normalized triangular functions. The rules comprise heuristic and fuzzy knowledge together with numerical data obtained from a geometric analysis of the control problem that considers the kinematic and dynamic constraints of the robot. Numerical data allow tuning the fuzzy symbols used in the rules to optimize the controller performance. From the implementation point of view, very few computational and memory resources are required: standard logical, addition, and multiplication operations and a few data that can be represented by integer values. This is illustrated with the design of a controller for the safe navigation of an autonomous car-like robot among possible obstacles toward a goal configuration. Implementation results of an FPGA embedded system based on a general-purpose soft processor confirm that percentage reduction in clock cycles is drastic thanks to applying the proposed neuro-fuzzy techniques. Simulation and experimental results obtained with the robot confirm the efficiency of the controller designed. Design methodology has been supported by the CAD tools of the environment Xfuzzy 3 and by the Embedded System Tools from Xilinx.     

Probabilistic models for robot-based object segmentation

This paper introduces a novel probabilistic method for robot based object segmentation. The method integrates knowledge of the robot’s motion to determine the shape and location of objects. This allows a robot with no prior knowledge of its workspace to isolate objects against their surroundings by moving them and observing their visual feedback. The main contribution of the paper is to improve upon current methods by allowing object segmentation in changing environments and moving backgrounds. The approach allows optimal values for the algorithm parameters to be estimated. Empirical studies against alternatives demonstrate clear improvements in both planar and three dimensional motion.     

Learning decision trees from decision rules: A method and initial results from a comparative study

A standard approach to determining decision trees is to learn them from examples. A disadvantage of this approach is that once a decision tree is learned, it is difficult to modify it to suit different decision making situations. Such problems arise, for example, when an attribute assigned to some node cannot be measured, or there is a significant change in the costs of measuring attributes or in the frequency distribution of events from different decision classes. An attractive approach to resolving this problem is to learn and store knowledge in the form of decision rules, and to generate from them, whenever needed, a decision tree that is most suitable in a given situation. An additional advantage of such an approach is that it facilitates buildingcompact decision trees, which can be much simpler than the logically equivalent conventional decision trees (by compact trees are meant decision trees that may contain branches assigned aset of values, and nodes assignedderived attributes, i.e., attributes that are logical or mathematical functions of the original ones). The paper describes an efficient method, AQDT-1, that takes decision rules generated by an AQ-type learning system (AQ15 or AQ17), and builds from them a decision tree optimizing a given optimality criterion. The method can work in two modes: thestandard mode, which produces conventional decision trees, andcompact mode, which produces compact decision trees. The preliminary experiments with AQDT-1 have shown that the decision trees generated by it from decision rules (conventional and compact) have outperformed those generated from examples by the well-known C4.5 program both in terms of their simplicity and their predictive accuracy.     

Using a Product Manifold distance for unsupervised action recognition

This paper presents a method for unsupervised learning and recognition of human actions in video. Lacking any supervision, there is nothing except the inherent biases of a given representation to guide grouping of video clips along semantically meaningful partitions. Thus, in the first part of this paper, we compare two contemporary methods, Bag of Features (BOF) and Product Manifolds (PM), for clustering video clips of human facial expressions, hand gestures, and full-body actions, with the goal of better understanding how well these very different approaches to behavior recognition produce semantically relevant clustering of data.     

强化学习在移动机器人自主导航中的应用

概述了移动机器人常用的自主导航算法及其优缺点,在此基础上提出了强化学习方法。描述了强化学习算法的原理,并实现了用神经网络解决泛化问题。设计了基于障碍物探测传感器信息的机器人自主导航强化学习方法,给出了学习算法中各要素的数学模型。经仿真验证,算法正确有效,具有良好的收敛性和泛化能力。     

机器人运动轨迹的模仿学习综述EI北大核心CSCD

作为机器人技能学习中的一个重要分支,模仿学习近年来在机器人系统中得到了广泛的应用.模仿学习能够将人类的技能以一种相对直接的方式迁移到机器人系统中,其思路是先从少量示教样本中提取相应的运动特征,然后将该特征泛化到新的情形.本文针对机器人运动轨迹的模仿学习进行综述.首先详细解释模仿学习中的技能泛化、收敛性和外插等基本问题;其次从原理上对动态运动基元、概率运动基元和核化运动基元等主要的模仿学习算法进行介绍;然后深入地讨论模仿学习中姿态和刚度矩阵的学习问题、协同和不确定性预测的问题以及人机交互中的模仿学习等若干关键问题;最后本文探讨了结合因果推理的模仿学习等几个未来的发展方向.     

Switching control approach for stable navigation of mobile robots in unknown environments

This paper presents a stable switching control strategy for the parking problem of non-holonomic mobile robots. First, it is proposed a positioning-orientation switching controller for the parking problem. With this strategy robot backwards motions are avoided and the robot heading is always in the direction of the goal point facilitating the obstacle handling. Second, the avoidance of unexpected obstacles is considered in a reactive way by following the contour of the obstacles. Next, the stability of the switching parking/obstacle-avoider controller is analyzed showing stability under reasonable conditions. Finally, the good performance and the feasibility of this approach are shown through several experimental results.