ICRA 2018 论文列表

Autonomous driving in urban areas requires avoiding other road users with only partial observability of the environment. Observations are only partial because obstacles can occlude the field of view of the sensors. The problem of robust and efficient navigation under uncertainty can be framed as a partially observable Markov decision process (POMDP). In order to bypass the computational cost of scaling the formulation to avoiding multiple road users, this paper demonstrates a decomposition method that leverages the optimal avoidance strategy for a single user. We evaluate the performance of two POMDP solution techniques augmented with the decomposition method for scenarios involving a pedestrian crosswalk and an intersection.

Within the European project DexROV the topic of underwater intervention is addressed. In particular, a remote control room is connected through a satellite communication link to surface vessel, which is in turn connected to an UVMS (Underwater Vehicle-Manipulator System) with an umbilical cable. The operator may interact with the system using a joystick or exoskeleton. Since a direct teleoperation is not feasible, a cognitive engine is in charge of handling communication latency or interruptions caused by the satellite link, and the UVMS should have sufficient autonomy in dealing with low level constraints or secondary objectives. To this purpose, a task-priority-based inverse kinematics algorithm has been developed in order to allow the operator to control only the end effector, while the algorithm is in charge of handling both operative and joint-space constraints. This paper describes some preliminary experimental results achieved during the DexROV campaign of July 2017 in Marseilles (France), where most of the components have been successfully integrated and the inverse kinematics nicely run.

While using temporal logic specifications with motion planning has been heavily researched, the reliance on having an available steering function is impractical and often suited only to basic problems with linear dynamics. This is because a steering function is a solution to an optimal two-point boundary value problem (OBVP); to our knowledge, it is nearly impossible to find an analytic solution to such problems in many cases. Addressing this issue, we have developed a means of combining the asymptotically optimal and probabilistically complete kinodynamic planning algorithm SST* with a local deterministic μ-calculus model checking procedure to create a motion planning algorithm with deterministic μ-calculus specifications that does not rely on a steering function. The procedure involves combining only the most pertinent information from multiple Kripke structures in order to create one abstracted Kripke structure storing the best paths to all possible proposition regions of the state-space. A linear-quadratic regulator (LQR) feedback control policy is then used to track these best paths, effectively connecting the trajectories found from multiple Kripke structures. Simulations demonstrate that it is possible to satisfy a complex liveness specification for infinitely often reaching specified regions of state-space using only forward propagation.

Search-and-rescue (SAR) robots operating in post-disaster urban areas need to accurately identify physical site information to perform navigation, mapping and manipulation tasks. This can be achieved by acquiring a 3D point cloud of the environment and performing object recognition from the point cloud data. However, this task is complicated by the unstructured environments and potentially-deformed objects encountered during disaster relief operations. Current 3D object recognition methods rely on point cloud input acquired under suitable conditions and do not consider deformations such as outlier noise, bending and truncation. This work introduces a deep learning architecture for 3D class recognition from point clouds of deformed building elements. The classification network, consisting of stacked convolution and average pooling layers applied directly to point coordinates, was trained using point clouds sampled from a database of mesh models. The proposed method achieves robustness to input variability using point sorting, resampling, and rotation normalization techniques. Experimental results on synthetically-deformed object datasets show that the proposed method outperforms the conventional deep learning methods in terms of classification accuracy and computational efficiency.

While reinforcement learning has led to promising results in robotics, defining an informative reward function is challenging. Prior work considered including the human in the loop to jointly learn the reward function and the optimal policy. Generating samples from a physical robot and requesting human feedback are both taxing efforts for which efficiency is critical. We propose to learn reward functions from both the robot and the human perspectives to improve on both efficiency metrics. Learning a reward function from the human perspective increases feedback efficiency by assuming that humans rank trajectories according to a low-dimensional outcome space. Learning a reward function from the robot perspective circumvents the need for a dynamics model while retaining the sample efficiency of model-based approaches. We provide an algorithm that incorporates bi-perspective reward learning into a general hierarchical reinforcement learning framework and demonstrate the merits of our approach on a toy task and a simulated robot grasping task.

In this paper, a teleoperation control of a 7-DoF robot manipulator for Minimally Invasive Surgery (MIS), which guarantees a safety-enhanced compliant behavior in the null space, is described. The redundancy of the manipulator is exploited to provide a flexible workspace for nurses or other staff (assisting physicians, patient support). The issue with safety and accurate surgical task execution may arise in the presence of human-robot interaction. Based on the implemented impedance control of tele-operated MIS tasks, a safety enhanced constraint is applied on the compliant null space motion. At the same time, the control approach integrates an adaptive fuzzy compensator to guarantee the accuracy of the surgical tasks during the uncertain human-robot interaction. The performance of the proposed algorithm is verified with virtual surgical tasks. The results showed that the compliant null space motion is constrained in a safe area, and also that the accuracy of tool tip is improved, providing a flexible and safe collaborative behavior in the null space for human-robot interaction during surgical tasks.

Safety is essential for robots in unknown environments, especially when there is physical Human-Robot Interaction (pHRI). Control over energy, or passivity, is an effective safety mechanism. However, when the control algorithm is implemented in a discrete-time computer, computation and communication delays readily lead to loss of passivity and to instability. In this paper, a way to make the actuators aware of the energy that they inject into the system is presented. Passivity and stability are then always guaranteed, even in situations of total communication loss. These Embedded Energy-Aware Actuators are a model-free passivity and safety layer that make complex robotic systems dependable, well-behaved and safe. The proposed method is validated in simulation and experiments.

A significant barrier regarding a successful implementation of fenceless robot cells into manufacturing areas with humans is given by the inefficiency due to safety requirements. Robot motions have to be slowed down so that an unexpected collision with a human does not result in human injuries. This velocity reduction leads to longer cycle times and, hence, fenceless robot cells turn out as uneconomic. In this paper, a new approach for human-robot collaboration in assembly tasks is presented. For a better performance of the robot, methods are investigated on how the robot can exploit a maximum performance while maintaining the safety of collaborating humans. For this purpose, the kinematics and dynamics of a human arm are described by a control-oriented dynamic model to determine its capability and reachability. Successful experiments validate the dynamic model as well as a corresponding projection approach for calculating possible movements of the human arm that may lead to a collision with the robot. Finally, this information is used to calculate an admissible path velocity that minimizes the danger of human injuries.

This paper presents a method for assisting human operators to teleoperate highly dynamic systems such as quadrotors inside a constrained environment with safety guarantees. Our method enables human operators to focus on manually operating and flying quadrotor systems without the need to focus on avoiding potential obstacles. This is achieved with the presented supervisory controller overriding human input to enforce safety constraints when necessary. This method can be used as an assistive training solution for novice pilots to begin flying quadrotors without crashing them. Our supervisory controller uses an Exponential control barrier function based quadratic program to achieve safe human teleoperated flight. We demonstrate and validate our control approach through several experiments with multiple users with varying skill levels for three different scenarios of a quadrotor flying in a motion capture environment with virtual and physical constraints.

To effectively control complex dynamical systems, accurate nonlinear models are typically needed. However, these models are not always known. In this paper, we present a data-driven approach based on Gaussian processes that learns models of quadrotors operating in partially unknown environments. What makes this challenging is that if the learning process is not carefully controlled, the system will go unstable, i.e., the quadcopter will crash. To this end, barrier certificates are employed for safe learning. The barrier certificates establish a non-conservative forward invariant safe region, in which high probability safety guarantees are provided based on the statistics of the Gaussian Process. A learning controller is designed to efficiently explore those uncertain states and expand the barrier certified safe region based on an adaptive sampling scheme. Simulation results are provided to demonstrate the effectiveness of the proposed approach.

This paper introduces the Buffered Input Cell as a reciprocal collision avoidance method for multiple vehicles with high-order linear dynamics, extending recently proposed methods based on the Buffered Voronoi Cell [1] and generalized Voronoi diagrams [2]. We prove that if each vehicle's control input remains in its Buffered Input Cell at each time step, collisions will be avoided indefinitely. The method is fast, reactive, and only requires that each vehicle measures the relative position of neighboring vehicles. We incorporate this collision avoidance method as one layer of a complete lane change control stack for autonomous cars in a freeway driving scenario. The lane change control stack comprises a decision-making layer, a trajectory planning layer, a trajectory following feedback controller, and the Buffered Input Cell for collision avoidance. We show in simulations that collisions are avoided with multiple vehicles simultaneously changing lanes on a freeway. We also show in simulations that autonomous cars using the BIC method effectively avoid collisions with an aggressive human-driven car.

SLAM is becoming a key component of robotics and augmented reality (AR) systems. While a large number of SLAM algorithms have been presented, there has been little effort to unify the interface of such algorithms, or to perform a holistic comparison of their capabilities. This is a problem since different SLAM applications can have different functional and non-functional requirements. For example, a mobile phone-based AR application has a tight energy budget, while a UAV navigation system usually requires high accuracy. SLAMBench2 is a benchmarking framework to evaluate existing and future SLAM systems, both open and close source, over an extensible list of datasets, while using a comparable and clearly specified list of performance metrics. A wide variety of existing SLAM algorithms and datasets is supported, e.g. ElasticFusion, InfiniTAM, ORB-SLAM2, OKVIS, and integrating new ones is straightforward and clearly specified by the framework. SLAMBench2 is a publicly-available software framework which represents a starting point for quantitative, comparable and val-idatable experimental research to investigate trade-offs across SLAM systems.

In this work, we present an approach to deep visuomotor control using structured deep dynamics models. Our model, a variant of SE3-Nets, learns a low-dimensional pose embedding for visuomotor control via an encoder-decoder structure. Unlike prior work, our model is structured: given an input scene, our network explicitly learns to segment salient parts and predict their pose embedding and motion, modeled as a change in the pose due to the applied actions. We train our model using a pair of point clouds separated by an action and show that given supervision only through point-wise data associations between the frames our network is able to learn a meaningful segmentation of the scene along with consistent poses. We further show that our model can be used for closed-loop control directly in the learned low-dimensional pose space, where the actions are computed by minimizing pose error using gradient-based methods, similar to traditional model-based control. We present results on controlling a Baxter robot from raw depth data in simulation and RGBD data in the real world and compare against two baseline deep networks. We also test the robustness and generalization performance of our controller under changes in camera pose, lighting, occlusion, and motion. Our method is robust, runs in real-time, achieves good prediction of scene dynamics, and outperforms baselines on multiple control runs. Video results can be found at: https://rse-lab.cs.washington.edu/se3-structured-deep-ctrl/.

We present a new efficient control framework for controlling groups of heterogeneous stress-engineered MEMS microrobots for accomplishing micro-assembly. The objective is to maximize the number of controllable microrobots in the system while keeping the number of external global signals as low as possible. This work proposes a theoretical control strategy that could complete multiple-shapes microassembly from arbitrary initial configuration where all the control primitives can be accompanied with a constant number (O(1)) of control pulses of the power delivery waveform. We focus on microrobotic systems that can be modeled as nonholonomic unicycles. We validate the control policy with hardware experiments for implementing planar assembly using multiple macroscale robots with direct drive wheels. These results lay the foundation for developing new methods to control of a large number of MEMS microrobots.

We consider the problem of guiding a senior user along a path using a robotic walking assistant. This is a particular type of path following problem, for which most of the solutions available in the literature require an exact localisation of the robot in the environment. An accurate localisation is obtained either with a heavy infrastructure (e.g., an active sensing system deployed in the environment or deploying landmarks in known positions) or using SLAM approaches with a massive data collection. Our key observation is that the intervention of the system (and a good level of accuracy) is only required in proximity of difficult decision points, while we can rely on the user in an environment where the only possibility is just to maintain a course (e.g., a corridor). The direct implication is that we can instrument the environment with a heavy infrastructure only in certain areas. This design strategy has to be complemented by an adequate control law that shifts the authority (i.e., the control of the actuators) between the robot and the user according to the accuracy of the information available to the robot. Such a control law is exactly the contribution of this paper.

When robots navigate through vineyards to perform irrigation adjustments, an optimization problem emerges whereby robots are tasked with performing adjustments having the highest cumulative outcome within a given temporal budget due to limited battery charge. To this end, the robot needs to reach a set of spatially distributed sites, and the specific structure of the vineyard imposes various constraints on possible motions. In this paper we first demonstrate that this type of orienteering problem remains NP-hard even for the restricted class of graphs associated with precision irrigation. Then, we devise and analyze two greedy heuristics informed by the problem we consider. Finally, these algorithms are evaluated on settings associated with a commercial vineyard and we show that our methods favorably compare to solutions proposed in the past.

Indoor plants in homes and commercial buildings such as malls, offices, airports, and hotels, can benefit from precision irrigation to maintain healthy growth and reduce water consumption. As active valves are too costly, and ongoing precise manual adjustment of drip emitters is impractical, we explore how the Toyota HSR mobile manipulator robot can autonomously adjust low-cost passive emitters. To provide sufficient accuracy for gripper alignment, we designed a lightweight, modular Emitter Localization Device (ELD) with cameras and LEDs that can be non-invasively mounted on the arm. This paper presents details of the design, algorithms, and experiments with adjusting emitters using a two-phase procedure: (1) aligning the robot base using the build-in hand camera, and (2) aligning the gripper axis with the emitter axis using the ELD. We report success rates and sensitivity analysis to tune computer vision parameters and joint motor gains. Experiments suggest that emitters can be adjusted with 95 % success rate in approximately 20 seconds.

In automated manufacturing, robots must reliably assemble parts of various geometries and low tolerances. Ideally, they plan the required motions autonomously. This poses a substantial challenge due to high-dimensional state spaces and non-linear contact-dynamics. Furthermore, object poses and model parameters, such as friction, are not exactly known and a source of uncertainty. The method proposed in this paper models the task of parts assembly as a belief space planning problem over an underlying impedance-controlled, compliant system. To solve this planning problem we introduce an asymptotically optimal belief space planner by extending an optimal, randomized, kinodynamic motion planner to nondeterministic domains. Under an expansiveness assumption we establish probabilistic completeness and asymptotic optimality. We validate our approach in thorough, simulated and realworld experiments of multiple assembly tasks. The experiments demonstrate our planner's ability to reliably assemble objects, solely based on CAD models as input.

We present a robust and precise localization system that achieves centimeter-level localization accuracy in disparate city scenes. Our system adaptively uses information from complementary sensors such as GNSS, LiDAR, and IMU to achieve high localization accuracy and resilience in challenging scenes, such as urban downtown, highways, and tunnels. Rather than relying only on LiDAR intensity or 3D geometry, we make innovative use of LiDAR intensity and altitude cues to significantly improve localization system accuracy and robustness. Our GNSS RTK module utilizes the help of the multi-sensor fusion framework and achieves a better ambiguity resolution success rate. An error-state Kalman filter is applied to fuse the localization measurements from different sources with novel uncertainty estimation. We validate, in detail, the effectiveness of our approaches, achieving 5-10cm RMS accuracy and outperforming previous state-of-the-art systems. Importantly, our system, while deployed in a large autonomous driving fleet, made our vehicles fully autonomous in crowded city streets despite road construction that occurred from time to time. A dataset including more than 60 km real traffic driving in various urban roads is used to comprehensively test our system.