ICRA 2020 论文列表

Structures with origami design enable objects to transform into various three-dimensional shapes. Traditionally origami structures are designed with zero-thickness flat paper sheets. However, the thickness and intersection of origami facets are non-negligible in most cases, uniquely when integrating origami design with robotic design because of the more efficient force transfer between thick plates compared with zero-thickness paper-sheets. Meanwhile, the single-layer-paper oriented initial design limited the shape transformation potential as multiple layer origami structures could conduct more variety of deformation. In this article, we are proposing a general design method of parallel-motion thick origami structures that could apply in robotic design like a parallel-motion gripper.

We propose a gripper displacement-magnification mechanism and an extendable finger mechanism, both of which can be attached to a commercially available parallel gripper. We then verify the operation of the mechanism in order to expand applications of the parallel gripper. The displacement-magnification mechanism has a stacked rack-and-pinion system that doubles displacement. The extendable finger mechanism has two nails that extend and contract, reducing impact force and detecting changes in product height from expansion and contraction amounts. The parallel gripper has a width of 95 mm, a depth of 110 mm, and a height of 214 mm and weighs 1.36 kg. It has an open/close stroke of 60 mm, a gripping force of 7.4 N, and an opening/closing speed of 100 mm/s or more. Further, it was confirmed that the ends and inclinations of products can be reliably detected using the extending/contracting nail. The mechanism verification confirmed that our parallel gripper achieved the desired performance and is therefore useful.

This paper presents a parallel permutation algorithm that achieves linear full-resolution reconfiguration of sliding-only cubic modular robots. We assume the use of a cubic module that can only slide across other modules' surfaces. The idea of a cubic modular robot with sliding-only motion primitive is a new concept that has advantages in simplifying the mechanisms of module hardware and space saving in its heterogeneous operations compared with previously studied cubic modules, such as those with sliding and convex motion primitives, or rotating motion primitives. However, because of its limited mobility, there are difficulties in managing the connectivity and scalability of the heterogeneous reconfiguration algorithm for it. To overcome these disadvantages, we introduce a parallel heterogeneous permutation method with linear operating time cost that can be incorporated into our previous full-resolution reconfiguration algorithm. We prove the correctness and completeness of the proposed algorithm. Simulation results show that the full-resolution reconfiguration algorithm that incorporates the proposed permutation algorithm reconfigures the robot structure with sliding-only cubic modules in linear operating-time cost.

In this work we propose long wave infrared (LWIR) imagery as a viable supporting modality for semantic segmentation using learning-based techniques. We first address the problem of RGB-thermal camera calibration by proposing a passive calibration target and procedure that is both portable and easy to use. Second, we present PST900, a dataset of 894 synchronized and calibrated RGB and Thermal image pairs with per pixel human annotations across four distinct classes from the DARPA Subterranean Challenge. Lastly, we propose a CNN architecture for fast semantic segmentation that combines both RGB and Thermal imagery in a way that leverages RGB imagery independently. We compare our method against the state-of-the-art and show that our method outperforms them in our dataset.

Mobile robot assistants have many applications, such as helping people in their daily living activities. These robots have to detect and recognize the actions and goals of the humans they are assisting. While there are several wide-spread plan and activity recognition solutions for controlled environments with many built-in sensors, like smart-homes, there is a lack of such systems for mobile robots operating in open settings, such as an apartment. We propose a module for the recognition of activities and goals for daily living by mobile robots, in real time and for complex activities. Our approach recognizes human-object interaction using an RGB-D camera to infer low-level actions which are sent to a goal recognition algorithm. Results show that our approach is both in real time and requires little computational resources, which facilitates its deployment on a mobile and low-cost robotics platform.

Flexible manufacturing and automation require robots that can be adapted to changing tasks. We propose to use modular robots that are customized from given modules for a specific task. This work presents an algorithm for proposing a module composition that is optimal with respect to performance metrics such as cycle time and energy efficiency, while considering kinematic, dynamic, and obstacle constraints. Tasks are defined as trajectories in Cartesian space, as a list of poses for the robot to reach as fast as possible, or as dexterity in a desired workspace. In a simulated comparison with commercially available industrial robots, we demonstrate the superiority of our approach in randomly generated tasks with respect to the chosen performance metrics. We use our modular robot proModular.1 for the comparison.

In this paper, we optimize the interaction graph of a multi-robot system (MRS) by maximizing its probability of security while requiring the MRS to have the fewest edges possible. Edges that represent robot interactions exist according to a probability distribution and security is defined using the control theoretic notion of left invertibility. To compute an optimal solution to our problem, we first start by reducing our problem to a variation of the rooted k-connections problem using three graph transformations. Then, we apply a weighted matroid intersection algorithm (WMIA) on matroids defined on the edge set of the interaction graph. Although the optimal solution can be found in polynomial time, MRSs are dynamic and their topologies may change faster than the rate at which the optimal security solution can be found. To cope with dynamic behavior, we present two heuristics that relax optimality but execute with much lower time complexity. Finally, we validate our results through Monte Carlo simulations.

Energy-efficient locomotion is of primary importance for legged robot to extend operation time in practical applications. This paper presents an approach to achieve energy-efficient locomotion for a quadrupedal robot walking over challenging terrains. Firstly, we optimize the nominal stance parameters based on the analysis of leg torque distribution. Secondly, we proposed the foothold planner and the center of gravity (COG) trajectory planner working together to guide the robot to place its standing legs in an energy-saving stance posture. We have validated the effectiveness of our method on a real quadrupedal robot in experiments including autonomously walking on plain ground and climbing stairs.

In this paper, we address the problem of optimal topology selection for stable coordination of multi-robot systems with asymmetric interactions. This problem arises naturally for multi-robot systems that interact based on sensing, e.g., with limited field of view (FOV) cameras. From our previous efforts on motion control in such settings, we have shown that not all interaction topologies yield stable coordinated motion when asymmetry exists. At the same time, not all robot-to-robot interactions are of equal quality, and thus we seek to optimize asymmetric interaction topologies subject to the constraint that the topology yields stable multi-robot motion. In this context, we formulate an optimal topology selection problem (OTSP) as a mixed integer semidefinite programming (MISDP) problem to compute optimal topologies that yield stable coordinated motion. Simulation results are provided to corroborate the effectiveness of the proposed OTSP formulation.

We study the problem of sequential task assignment and collision-free routing for large teams of robots in applications with inter-task precedence constraints (e.g., task A and task B must both be completed before task C may begin). Such problems commonly occur in assembly planning for robotic manufacturing applications, in which sub-assemblies must be completed before they can be combined to form the final product. We propose a hierarchical algorithm for computing makespan-optimal solutions to the problem. The algorithm is evaluated on a set of randomly generated problem instances where robots must transport objects between stations in a "factory" grid world environment. In addition, we demonstrate in high-fidelity simulation that the output of our algorithm can be used to generate collision-free trajectories for non-holonomic differential-drive robots.

In this paper, we consider the Constant-cost Orienteering Problem (COP) where a robot, constrained by a limited travel budget, aims at selecting a path with the largest reward in an aisle-graph. The aisle-graph consists of a set of loosely connected rows where the robot can change lane only at either end, but not in the middle. Even when considering this special type of graphs, the orienteering problem is known to be intractable. We optimally solve in polynomial time two special cases, COP-FR where the robot can only traverse full rows, and COP-SC where the robot can access the rows only from one side. To solve the general COP, we then apply our special case algorithms as well as a new heuristic that suitably combines them. Despite its light computational complexity and being confined into a very limited class of paths, the optimal solutions for COP-FR turn out to be competitive in terms of achieved rewards even for COP. This is shown by means of extended simulations performed on both real and synthetic scenarios. Furthermore, our new heuristic for the general case outperforms state-of-art algorithms, especially for input with highly unbalanced rewards.

State-of-the-art approaches to legged locomotion are widely dependent on the use of models like the linear inverted pendulum (LIP) and the spring-loaded inverted pendulum (SLIP), popular because their simplicity enables a wide array of tools for planning, control, and analysis. However, they inevitably limit the ability to execute complex tasks or agile maneuvers. In this work, we aim to automatically synthesize models that remain low-dimensional but retain the capabilities of the high-dimensional system. For example, if one were to restore a small degree of complexity to LIP, SLIP, or a similar model, our approach discovers the form of that additional complexity which optimizes performance. In this paper, we define a class of reduced-order models and provide an algorithm for optimization within this class. To demonstrate our method, we optimize models for walking at a range of speeds and ground inclines, for both a five-link model and the Cassie bipedal robot.

This paper presents a design of oscillation damping control for the cable-Suspended Aerial Manipulator (SAM). The SAM is modeled as a double pendulum, and it can generate a body wrench as a control action. The main challenge is the fact that there is only one onboard IMU sensor which does not provide full information on the system state. To overcome this difficulty, we design a controller motivated by a simplified SAM model. The proposed controller is very simple yet robust to model uncertainties. Moreover, we propose a gain tuning rule by formulating the proposed controller in the form of output feedback linear quadratic regulation problem. Consequently, it is possible to quickly dampen oscillations with minimal energy consumption. The proposed approach is validated through simulations and experiments.

After occasional perturbation, it is crucial to spontaneously control the limit cycle walking so that it quickly returns to its closed orbit in phase space. Otherwise, its stability can not be sufficiently guaranteed if the speed of recovery is slow while successive perturbation is applied. The accumulated deviation may eventually drive the phase outside the basin of attraction, leading to failure of the walking. In this sense, a control law that quickly recovers the disturbed phase before encountering the following perturbations is indispensable. With this consideration, here we analytically derive an optimal fast entrainment waveform that maximizes the speed of phase recovery based on phase reduction theory. Our theoretical method is numerically evaluated using a limit cycle walker, which is indirectly controlled by the oscillation of a wobbling mass via entrainment effect. The obtained waveform is used as the desired trajectory of the wobbling motion. The simulation results show that the waveform we derived achieves the best performance among all candidates. Our method helps to enhance the stability of limit cycle walking.

Actuator power consumption is a limiting factor in mobile robot design. In this paper we introduce the concept of an energy-recycling actuator, which uses an array of springs and clutches to capture and return elastic energy in parallel with an electric motor. Engaging and disengaging clutches appropriately could reduce electrical energy consumption without sacrificing controllability, but presents a challenging control problem. We formulated the optimal control objective of minimizing actuator power consumption as a mixed-integer quadratic program (MIQP) and solved for the global minimum. For a given actuator design and a wide range of simulated torque and rotation patterns, all corresponding to zero net work over one cycle, we compared optimized actuator energy consumption to that of an optimized gear motor with simple parallel elasticity. The simulated energy-recycling actuator consumed less electrical energy: 57% less on average and 80% less in the best case. These results demonstrate an effective approach to optimal control of this type of system, and suggest that energy-recycling actuators could substantially reduce power consumption in some robotics applications.

In this paper, we present an open platform, termed OpenVINS, for visual-inertial estimation research for both the academic community and practitioners from industry. The open sourced codebase provides a foundation for researchers and engineers to quickly start developing new capabilities for their visual-inertial systems. This codebase has out of the box support for commonly desired visual-inertial estimation features, which include: (i) on-manifold sliding window Kalman filter, (ii) online camera intrinsic and extrinsic calibration, (iii) camera to inertial sensor time offset calibration, (iv) SLAM landmarks with different representations and consistent First-Estimates Jacobian (FEJ) treatments, (v) modular type system for state management, (vi) extendable visual-inertial system simulator, and (vii) extensive toolbox for algorithm evaluation. Moreover, we have also focused on detailed documentation and theoretical derivations to support rapid development and research, which are greatly lacked in the current open sourced algorithms. Finally, we perform comprehensive validation of the proposed OpenVINS against state-of-the-art open sourced algorithms, showing its competing estimation performance.

The recent breakthroughs in computer vision have benefited from the availability of large representative datasets (e.g. ImageNet and COCO) for training. Yet, robotic vision poses unique challenges for applying visual algorithms developed from these standard computer vision datasets due to their implicit assumption over non-varying distributions for a fixed set of tasks. Fully retraining models each time a new task becomes available is infeasible due to computational, storage and sometimes privacy issues, while naïve incremental strategies have been shown to suffer from catastrophic forgetting. It is crucial for the robots to operate continuously under open-set and detrimental conditions with adaptive visual perceptual systems, where lifelong learning is a fundamental capability. However, very few datasets and benchmarks are available to evaluate and compare emerging techniques. To fill this gap, we provide a new lifelong robotic vision dataset ("OpenLORIS-Object") collected via RGB-D cameras. The dataset embeds the challenges faced by a robot in the real-life application and provides new benchmarks for validating lifelong object recognition algorithms. Moreover, we have provided a testbed of 9 state-of-the-art lifelong learning algorithms. Each of them involves 48 tasks with 4 evaluation metrics over the OpenLORIS-Object dataset. The results demonstrate that the object recognition task in the ever-changing difficulty environments is far from being solved and the bottlenecks are at the forward/backward transfer designs. Our dataset and benchmark are publicly available at https://lifelong-robotic-vision.github.io/dataset/object.

With new, safer manipulator robots, the probability of serious injury due to collisions with humans remains low (5%), even at speeds as high as 2 m.s-1. Collisions would better be avoided nevertheless, because they disrupt the tasks of both the robot and the human. We propose in this paper to equip robots with exteroceptive sensors and online motion generation so that the robot is able to perceive and react to the motion of the human in order to reduce the occurrence of collisions. It's impossible to guarantee that no collision will ever take place in a partially unknown dynamic environment such as a shared workspace, but we can guarantee instead that, if a collision takes place, the robot is at rest at the time of collision, so that it doesn't inject its own kinetic energy in the collision. To do so, we adapt a Model Predictive Control scheme which has been demonstrated previously with two industrial manipulator robots avoiding collisions while sharing their workspace. The proposed control scheme is validated in simulation.

The increasing availability of surround-view camera systems in passenger vehicles motivates their use as an exterior perception modality for intelligent vehicle behaviour. An important problem within this context is the extrinsic calibration between the cameras, which is challenging due to the often reduced overlap between the fields of view of neighbouring views. Our work is motivated by two insights. First, we argue that the accuracy of vision-based vehicle motion estimation depends crucially on the quality of exterior orientation calibration, while design parameters for camera positions typically provide sufficient accuracy. Second, we demonstrate how planar vehicle motion related direction vectors can be used to accurately identify individual camera-to-vehicle rotations, which are more useful than the commonly and tediously derived camera-to-camera transformations. We present a complete and highly practicable online optimisation strategy to obtain the exterior orientation parameters and conclude with successful tests on simulated, indoor, and large-scale outdoor experiments.