IROS 2019 论文列表

Fast and precise motion control is important for industrial robots in manufacturing applications. However, some collaborative robots sacrifice precision for safety, particular for high motion speed. The performance degradation is caused by the inability of the joint servo controller to address the uncertain nonlinear dynamics of the robot arm, e.g., due to joint flexibility. We consider two approaches to improve the trajectory tracking performance through feedforward compensation. The first approach uses iterative learning control, with the gradient-based iterative update generated from the robot forward dynamics model. The second approach uses dynamic inversion to directly compensate for the robot forward dynamics. If the forward dynamics is strictly proper or is non-minimum-phase (e.g., due to time delays), its stable inverse would be non-causal. Both approaches require robot dynamical models. This paper presents results of using recurrent neural networks (RNNs) to approximate these dynamical models - forward dynamics in the first case, inverse dynamics (possibly non-causal) in the second case. We use the bi-directional RNN to capture the noncausality. The RNNs are trained based on a collection of commanded trajectories and the actual robot responses. We use a Baxter robot to evaluate the two approaches. The Baxter robot exhibits significant joint flexibility due to the series-elastic joint actuators. Both approaches achieve sizable improvement over the uncompensated robot motion, for both random joint trajectories and Cartesian motion. The inverse dynamics method is particularly attractive as it may be used to more accurately track a user input as in teleoperation.

Fast and efficient path generation is critical for robots operating in complex environments. This motion planning problem is often performed in a robot's actuation or configuration space, where popular pathfinding methods such as A*, RRT*, get exponentially more computationally expensive to execute as the dimensionality increases or the spaces become more cluttered and complex. On the other hand, if one were to save the entire set of paths connecting all pair of locations in the configuration space a priori, one would run out of memory very quickly. In this work, we introduce a novel way of producing fast and optimal motion plans for static environments by using a stepping neural network approach, called OracleNet. OracleNet uses Recurrent Neural Networks to determine end-to-end trajectories in an iterative manner that implicitly generates optimal motion plans with minimal loss in performance in a compact form. The algorithm is straightforward in implementation while consistently generating near-optimal paths in a single, iterative, end-to-end roll-out. In practice, OracleNet generally has fixed-time execution regardless of the configuration space complexity while outperforming popular pathfinding algorithms in complex environments and higher dimensions1.

This paper presents a neural network based heterogeneous sensor fusion approach towards real-time traversability estimation of mobile robots using sensor data. Even though significant advances have been made for autonomous navigation in structured terrain conditions, obtaining reliable traversability estimates for tracked vehicle navigation in challenging terrain conditions is still an open research problem. In this regard, we propose a neural network architecture capable of fusing depth images along with roll and pitch measurements on board the robot to perform traversability estimation. The proposed architecture is trained in a variety of simulated structured and unstructured environments. As such, the proposed architecture is capable of extracting the relevant features from the sensor measurements in a data driven manner as compared to existing heuristic based approaches that fail to generalize for different environmental conditions. The reliability of the traversability estimates provided by the trained architecture was validated in indoor and outdoor conditions using real sensor data. In addition, the feasibility of using the traversability estimates in incremental path planning was also demonstrated through simulation. For both applications the proposed approach provided compelling results. Inferences based on the results of the experiments along with directions for future research are also outlined.

In order to use walking robots for exploration in a real complex environment, an adaptive control system is required to allow them to successfully and efficiently traverse the terrains. To achieve this, we propose here our adaptive locomotion control technique of a walking robot. It is based on a modular structure, combining neural control with an artificial hormone system. The neural control coordinates all leg joints of the robot and generates its locomotion with various insect-like gaits. In parallel, the artificial hormone system uses the motor commands from the neural control and foot contact feedback to estimate the walking state and automatically adapt the joint movements with respect to the terrain. The adaptability is quickly achieved in an online manner within a few seconds. Robot walking experiments show that this adaptive control technique enables a six-legged robot to adapt to various difficult terrains with energy efficiency. Such terrains include sand (loose ground), sponge with different softness levels (soft/compliant ground), grass (vegetated ground), and floor/pavement (hard ground). The technique does not require robot kinematics or an environmental model; and can, therefore, be potentially applied to different legged robots to achieve stable online adaptation and energy-efficient locomotion on unpredictable (compliant) terrains.

Grasping a simple object from the side is easy — unless the object is almost as big as the hand or space constraints require positioning the robot hand awkwardly with respect to the object. We show that humans — when faced with this challenge — adopt coordinated finger movements which enable them to successfully grasp objects even from these awkward poses. We also show that it is relatively straight forward to implement these strategies autonomously. Our human-studies approach asks participants to perform grasping task by either “puppetteering” a robotic manipulator that is identical (geometrically and kinematically) to a popular underactuated robotic manipulator (the Barrett hand), or using sliders to control the original Barrett hand. Unlike previous studies, this enables us to directly capture and compare human manipulation strategies with robotic ones. Our observation is that, while humans employ underactuation, how they use it is fundamentally different (and more effective) than that found in existing hardware.

Assistive robots can potentially improve the quality of life and personal independence of elderly people by supporting everyday life activities. To guarantee a safe and intuitive interaction between human and robot, human intentions need to be recognized automatically. As humans communicate their intentions multimodally, the use of multiple modalities for intention recognition may not just increase the robustness against failure of individual modalities but especially reduce the uncertainty about the intention to be recognized. This is desirable as particularly in direct interaction between robots and potentially vulnerable humans a minimal uncertainty about the situation as well as knowledge about this actual uncertainty is necessary. Thus, in contrast to existing methods, in this work a new approach for multimodal intention recognition is introduced that focuses on uncertainty reduction through classifier fusion. For the four considered modalities speech, gestures, gaze directions and scene objects individual intention classifiers are trained, all of which output a probability distribution over all possible intentions. By combining these output distributions using the Bayesian method Independent Opinion Pool [1] the uncertainty about the intention to be recognized can be decreased. The approach is evaluated in a collaborative human-robot interaction task with a 7-DoF robot arm. The results show that fused classifiers, which combine multiple modalities, outperform the respective individual base classifiers with respect to increased accuracy, robustness, and reduced uncertainty.

Generating robot motions from a precomputed graph has proven to be an effective approach to solving many motion planning problems. After their generation, roadmaps reduce complex motion planning problems to that of solving a graph-based shortest path. However, generating the graph can involve tradeoffs, such as how sparse or dense to make the graph. Sparse graphs may not provide enough options to navigate around a new obstacle or may result in grossly suboptimal motions. Dense graphs may take too long to search and result in an unresponsive robot. In this paper we present an algorithm that generates a graph with multiple sparse levels- the sparsest level can be searched quickly, while the densest level allows for asymptotically optimal motions. With the paired multilevel shortest path algorithm, after the robot computes an initial solution, it can then incrementally refine the shortest-path as time allows. We demonstrate the algorithms on an articulated robot with 8 degrees of freedom, having them compute an initial solution in a fraction of the time required for a full graph search, and subsequently, incrementally refine the solution to the optimal shortest path from the densest level of the graph.

Deformable objects and surfaces are ubiquitous in the daily lives of humans - from the garments in fashion to soft tissues within the body. Because of this routine interaction with soft materials, humans are adept and trained in manipulation of deformable objects while avoiding irreversible damage. The dexterity and care involved is largely facilitated through a combination of the human haptic sense of touch and visual observations of object deformation [1]. While this scenario presents itself as a trivially intuitive task, it becomes significantly more difficult and complex with the deprivation of both 3D depth perception and haptic senses. This deprived state is not dissimilar to the scenarios encountered in many robot-assisted minimally invasive surgeries. As a result, unintentional tissue damage can occur due to lack of force feedback and fine 3D visibility [2]. One approach to remediate these issues combines real-time dynamic 3D reconstruction and vision-based force estimation for haptic feedback. Toward that end, this work continues research in a series of studies focusing on multicamera 3D reconstruction of dynamic surgical cavities. Previous work introduced a novel approach of camera grouping and pair sequencing [3]. This paper builds upon that work by introducing a method for non-rigid, sparse point cloud registration and subsequent point classification. In particular, to enable deformation and force analyses, surfaces are locally classified into three categories: static, shifting and deforming. The topics addressed in this paper present open challenges and ongoing research directions for researchers to this day [4], and provide a step towards real-time 3D reconstruction and force feedback in robot-assisted surgery.

This paper presents an object-centric method for efficiently performing two types of challenging pick-and-place tasks, namely sequential pick and place and object sorting. We propose multiclass dense object nets (MCDONs) for learning object-centric dense descriptors that maintain not only intra-class variations but also inter-class separation. Intra-class consistency is also inherently learned and is useful for our pick-and-place tasks. All the tasks only require a single demonstration from users, which can then be generalized to all class instances. A dataset containing eight classes and a total of 52 objects was provided in this study. We obtained a task success rate of 93.33% on a five-block stacking task and 97.41% on a three-class object sorting task.

Different legged robot locomotion controllers offer different advantages; from speed of motion to energy, computational demand, safety and others. In this paper we propose a method for planning locomotion with multiple controllers and sub-planners, explicitly considering the multi-objective nature of the legged locomotion planning problem. The planner first obtains body paths extended with a choice of controller or sub-planner, and then fills the gaps by sub-planning. The method leads to paths with a mix of static and dynamic walking which only plan footsteps where necessary. We show that our approach is faster than pure footstep planning methods both in computation (2x) and mission time (1.4x), and safer than pure dynamic-walking methods. In addition, we propose two methods for aggregating the multiple objectives in search-based planning and reach desirable trade-offs without weight tuning. We show that they reach desirable Pareto-optimal solutions up to 8x faster than fairly-tuned traditional weighted-sum methods. Our conclusions are drawn from a combination of planning, physics simulation, and real robot experiments.

Local mapping plays an important role in outdoor intelligent vehicle applications and multi-vehicle cooperative local mapping which takes advantage of vehicular communication can bring considerable benefits to this important task. In this paper, a multi-vehicle cooperative local mapping architecture using split covariance intersection filter (Split CIF) is proposed. In the proposed method, a vehicle can flexibly perform cooperative local mapping with other vehicles in decentralized way, without complicated monitoring and controlling of data flow among vehicles; fused maps can be shared freely among vehicles. An efficient and accurate implementation of the Split CIF is also introduced. A simulation-based comparative study demonstrates the potential and advantage of the proposed multi-vehicle cooperative local mapping architecture using Split CIF.

The need for robots autonomously navigating in more and more complex environments has motivated intense R& D efforts in making robot pose estimation more accurate and reliable. This paper presents a multi-sensor multi-hypothesis method for robust 6-DoF localization in complex environments. Robustness and accuracy requirements are addressed as follows. First, camera and LIDAR features are seamlessly integrated in the same statistical framework, benefiting from their synergies and providing robustness in scenarios with low or varying densities of LIDAR and visual features. Second, a multi-hypothesis approach is adopted to cope with scenario symmetries. The method has been carefully designed to operate in real time using feature and hypothesis filtering and efficient hypothesis refinement, and has been coded in a multi-core implementation. The proposed method has been extensively validated for closed-loop aerial robot navigation in different urban and industrial scenarios and has shown advantages over well-known single-sensor techniques.

Multi-robot assembly has the potential to greatly reduce the cost and risk associated with the fabrication of large structures. Using teams of robots to perform assembly offers numerous advantages such as parallelism, robustness to single-agent failures, and flexibility in scheduling and task assignment. However, while previous work on multi-robot assembly focuses on generating feasible assembly plans and decentralized control strategies, we instead study the problem of planning optimal assembly sequences.To this end, we pose the problem of multi-robot assembly as a discrete optimization, specifically an integer linear program (ILP) or quadratic program (IQP), which aims to minimize the time to complete the assembly, or to minimize the distance traveled. We develop a model of multi-robot assembly that captures both geometric constraints and actuation constraints inherent to the problem. While the ILP and IQP can be solved exactly using commercial optimization software in a substantial amount of time, we also propose heuristic strategies which can be quickly computed, and can scale to structures of reasonable size. We also verify our methods empirically by comparing their performance on a variety of test structures.

In this paper, we propose a novel system architecture called multi-layer environmental affordance map for social and service companion robots. Based on this architecture, robots can organize the perception and inference information efficiently and generate social friendly navigation strategies. In other words, robots are able to strengthen their perception and inference abilities to interact with domestic environment and users under our efficient framework. The main feature of this architecture is that the relations between layers can be viewed as affordances to improve the accuracy and the robustness of the detection and inference. The results show that our architecture achieves robust indoor localization, scene localization, human event detection and socially friendly navigation in real time under limited computational resource.

We present a method to manipulate virtual objects which are constrained complexly and reconfigurable as if they would exist in a real world by using multiple hands simultaneously. A complexly constrained, and reconfigurable object, such as a Rubik's cube, is hard to describe its physical motion constraints, mainly because they are determined by the grasping situation and dynamically changeable. Rather than describing the physical motion constraints in a general form, we more focus on the multiple hand interaction of a complex object. A complex object is divided into multiple subparts which are grasped by each hand, and the constraints between the subparts are optimized for inducing natural and continuous movement. For this, we propose a dynamically adjustable data structure for representing object parts grasped by multiple hands, and an optimization-based pose estimation of the constrained subparts along with their grasped hands. The experiments show that human subjects can manipulate a complexly constrained object such as a Rubik's cube without any difficulty as if it exists in the real-world.

This paper focuses on a stabilization control for multi-contact motion which enables a humanoid robot to locomote by realizing dynamic contact transitions on non-flat environment. In the stabilization process of the multi-contact motion, the desired Zero-Moment Point (ZMP) is modified by the position of the Divergent Component of Motion (DCM) error with respect to the 3D Center of Mass (CoM) motion generated from the force distribution ratio. The contact wrench of each end-effector is determined by quadratic optimization considering the centroidal dynamics and contact friction constraints so as to satisfy the modified ZMP. Each end-effector is controlled by optimized force reference through a projection of null space by force distribution ratio. We propose a multi-contact stabilization framework which can be designed not only to generate 3D CoM motion but also the CoM position estimation and the optimal force distribution around the reference ZMP in a unified manner from a balance controller, by using the force distribution ratio. The effectiveness of proposed method is validated by a quadruped locomotion leaning against a vertical wall using the joint position controlled humanoid HRP-5P in a dynamic simulator.

We investigate a classification problem using multiple mobile agents capable of collecting (partial) pose-dependent observations of an unknown environment. The objective is to classify an image over a finite time horizon. We propose a network architecture on how agents should form a local belief, take local actions, and extract relevant features from their raw partial observations. Agents are allowed to exchange information with their neighboring agents to update their own beliefs. It is shown how reinforcement learning techniques can be utilized to achieve decentralized implementation of the classification problem by running a decentralized consensus protocol. Our experimental results on the MNIST handwritten digit dataset demonstrates the effectiveness of our proposed framework.

Autonomous vehicles are, in contrast to classical automated guided vehicles (AGVs), less predictable in their behavior and drive time. Therefore, the issue of how to efficiently control these vehicles arises, because autonomous agents need to be coordinated and not controlled, to give autonomous behaviors and actions space. The scientific contribution of this paper is a novel approach, based on prioritized planning to target this issue as well as an open source framework for evaluation and comparison. Prioritized planning has the disadvantage of being neither optimal nor complete, however, it has the advantage of being computationally feasible. This work utilizes prioritized planning to significantly increase the set of feasible scenarios through collision prevention: by locally finding alternative routes and adding them to the search graph. The paper clearly formulates the extensions needed and delineates the approach’s limits, as it is neither optimal nor complete. More importantly, however, our method calculates routes for each vehicle with inter-vehicle synchronization, enabling vehicles to execute the plan in a distributed fashion without centralized control, thereby allowing autonomous behavior. Finally, results are verified by comparing our Multi Robot Router (MRR) proposed in this work to classical approaches. The software developed as well as the test sets are publicly available for ROS and the simulation environment.

Recursive state estimation is often used to estimate a probability density function of a specific state, e.g. a robot's pose, over time. Compared to Kalman filters, Sequential Monte Carlo (SMC) methods are less constrained in regard to state propagation and update model definition, which makes it easier to implement any suitable problem. In this work, we present a generic Sequential Monte Carlo framework, which uses abstract formulations for importance weighting, propagation and resampling and provides an independent core algorithm that is usable for any problem instantiation, such that diverse SMC problems can be implemented easily and quickly, since the basic algorithms are already provided. Current applications include 2D localization, 2D tracking in a SLAM system and contact point localization on a manipulator surface. Further, we introduce concepts to deal with data input synchronization and fair execution of different weighting models, which makes it possible to incorporate data from as many update sources, e.g. sensors, as desired. As a typical application scenario, we provide a plugin-based and hence easily extensible instantiation for 2D localization and demonstrate the capabilities of our framework and methods based on a well-known dataset.