ICRA 2020 论文列表

Autonomous robots require online trajectory planning capability to operate in the real world. Efficient offline trajectory planning methods already exist, but are computationally demanding, preventing their use online. In this paper, we present a novel algorithm called Guided Trajectory Learning that learns a function approximation of solutions computed through trajectory optimization while ensuring accurate and reliable predictions. This function approximation is then used online to generate trajectories. This algorithm is designed to be easy to implement, and practical since it does not require massive computing power. It is readily applicable to any robotics systems and effortless to set up on real hardware since robust control strategies are usually already available. We demonstrate the computational performance of our algorithm on flat-foot walking with the self-balanced exoskeleton Atalante.

To solve multi-step manipulation tasks in the real world, an autonomous robot must take actions to observe its environment and react to unexpected observations. This may require opening a drawer to observe its contents or moving an object out of the way to examine the space behind it. Upon receiving a new observation, the robot must update its belief about the world and compute a new plan of action. In this work, we present an online planning and execution system for robots faced with these challenges. We perform deterministic cost-sensitive planning in the space of hybrid belief states to select likely-to-succeed observation actions and continuous control actions. After execution and observation, we replan using our new state estimate. We initially enforce that planner reuses the structure of the unexecuted tail of the last plan. This both improves planning efficiency and ensures that the overall policy does not undo its progress towards achieving the goal. Our approach is able to efficiently solve partially observable problems both in simulation and in a real-world kitchen.

We consider the kinodynamic multi-robot planning problem in cluttered 3-D workspaces. Reachability analysis on position invariant geometric trees is leveraged to find kino- dynamically feasible trajectories for the multi-robot team from potentially non-stationary initial states. The key contribution of our approach is that a collision-free geometric solution guarantees a kinodynamically feasible, safe solution without additional refinement. Simulation results with up-to 40 robots and hardware results with 5 robots suggest the viability of the proposed approach for online planning and replanning for large teams of aerial robots in cluttered 3-D workspaces.

In this paper, we present a 3D factor-graph LiDAR-SLAM system which incorporates a state-of-the-art deeply learned feature-based loop closure detector to enable a legged robot to localize and map in industrial environments. Point clouds are accumulated using an inertial-kinematic state estimator before being aligned using ICP registration. To close loops we use a loop proposal mechanism which matches individual segments between clouds. We trained a descriptor offline to match these segments. The efficiency of our method comes from carefully designing the network architecture to minimize the number of parameters such that this deep learning method can be deployed in real-time using only the CPU of a legged robot, a major contribution of this work. The set of odometry and loop closure factors are updated using pose graph optimization. Finally we present an efficient risk alignment prediction method which verifies the reliability of the registrations. Experimental results at an industrial facility demonstrated the robustness and flexibility of our system, including autonomous following paths derived from the SLAM map.

We propose a self-supervised approach for learning representations of objects from monocular videos and demonstrate it is particularly useful for robotics. The main contributions of this paper are: 1) a self-supervised model called Object-Contrastive Network (OCN) that can discover and disentangle object attributes from video without using any labels; 2) we leverage self-supervision for online adaptation: the longer our online model looks at objects in a video, the lower the object identification error, while the offline baseline remains with a large fixed error; 3) we show the usefulness of our approach for a robotic pointing task; a robot can point to objects similar to the one presented in front of it. Videos illustrating online object adaptation and robotic pointing are provided as supplementary material.

High-performance composites are increasingly being used in the industry. Sheet layup is a process of manufacturing composite components using deformable sheets. We have developed a robotic cell to automate the layup process and overcome the limitations of the manual layup. Generating offline trajectories for robots and executing them without online refinement can introduce defects in the process due to uncertainties in the model of the sheet and environmental factors. Our system computes layup and grasping trajectories for the robots and refines them during the layup process based on the sensor data. We use an approach that augments physical experiments with simulations to train a Gaussian process regression model offline. The use of GPR enables us to quickly refine grasp plans and perform a defect-free layup without slowing down the layup process. We present experimental results on two components.

As a crucial step of sensor data fusion, sensor calibration plays a vital role in many cutting-edge machine vision applications, such as autonomous vehicles and AR/VR. Existing techniques either require quite amount of manual work and complex settings, or are unrobust and prone to produce suboptimal results. In this paper, we investigate the extrinsic calibration of an RGB camera and a light detection and ranging (LiDAR) sensor, which are two of the most widely used sensors in autonomous vehicles for perceiving the outdoor environment. Specifically, we introduce an online calibration technique that automatically computes the optimal rigid motion transformation between the aforementioned two sensors and maximizes their mutual information of perceived data, without the need of tuning environment settings. By formulating the calibration as an optimization problem with a novel calibration quality metric based on semantic features, we successfully and robustly align pairs of temporally synchronized camera and LiDAR frames in real time. Demonstrated on several autonomous driving tasks, our method outperforms state-of-the-art edge feature based auto-calibration approaches in terms of robustness and accuracy.

Globally localizing in a given map is a crucial ability for robots to perform a wide range of autonomous navigation tasks. This paper presents OneShot - a global localization algorithm that uses only a single 3D LiDAR scan at a time, while outperforming approaches based on integrating a sequence of point clouds. Our approach, which does not require the robot to move, relies on learning-based descriptors of point cloud segments and computes the full 6 degree-of-freedom pose in a map. The segments are extracted from the current LiDAR scan and are matched against a database using the computed descriptors. Candidate matches are then verified with a geometric consistency test. We additionally present a strategy to further improve the performance of the segment descriptors by augmenting them with visual information provided by a camera. For this purpose, a custom-tailored neural network architecture is proposed. We demonstrate that our LiDAR-only approach outperforms a state-of-the-art baseline on a sequence of the KITTI dataset and also evaluate its performance on the challenging NCLT dataset. Finally, we show that fusing in visual information boosts segment retrieval rates by up to 26% compared to LiDAR-only description.

Path planning is important for robot action execution, since a path or a motion trajectory for a particular action has to be defined first before the action can be executed. Most of the current approaches are iterative methods where the trajectory is generated by predicting the next state based on the current state. Here we propose a novel method by utilising a fully convolutional neural network, which allows generation of complete paths even for several agents with one network prediction iteration. We demonstrate that our method is able to successfully generate optimal or close to optimal paths (less than 10% longer) in more than 99% of the cases for single path predictions in 2D and 3D environments. Furthermore, we show that the network is - without specific training on such cases - able to create (close to) optimal paths in 96% of the cases for two and in 84% of the cases for three simultaneously generated paths.

With the increase in use of Unmanned Aerial Vehicles (UAVs)/drones, it is important to detect and identify causes of failure in real time for proper recovery from a potential crash-like scenario or post incident forensics analysis. The cause of crash could be either a fault in the sensor/actuator system, a physical damage/attack, or a cyber attack on the drone's software. In this paper, we propose novel architectures based on deep Convolutional and Long Short-Term Memory Neural Networks (CNNs and LSTMs) to detect (via Autoencoder) and classify drone mis-operations based on real-time sensor data. The proposed architectures are able to learn high-level features automatically from the raw sensor data and learn the spatial and temporal dynamics in the sensor data. We validate the proposed deep-learning architectures via simulations and realworld experiments on a drone. Empirical results show that our solution is able to detect (with over 90% accuracy) and classify various types of drone mis-operations (with about 99% accuracy (simulation data) and upto 85% accuracy (experimental data)).

By leveraging our relative Micro-scale Unmanned Aerial Vehicle localization sensor UVDAR, we generated an automatically annotated dataset MIDGARD, which the community is invited to use for training and testing their machine learning systems for the detection and localization of Microscale Unmanned Aerial Vehicles (MAVs) by other MAVs. Furthermore, we provide our system as a mechanism for rapidly generating custom annotated datasets specifically tailored for the needs of a given application. The recent literature is rich in applications of machine learning methods in automation and robotics. One particular subset of these methods is visual object detection and localization, using means such as Convolutional Neural Networks, which nowadays enable objects to be detected and classified with previously inconceivable precision and reliability. Most of these applications, however, rely on a carefully crafted training dataset of annotated camera footage. These must contain the objects of interest in environments similar to those where the detector is expected to operate. Notably, the positions of the objects must be provided in annotations. For non-laboratory settings, the construction of such datasets requires many man-hours of manual annotation, which is especially the case for use onboard Micro-scale Unmanned Aerial Vehicles. In this paper, we are providing for the community a practical alternative to that kind of approach.

Torque distribution in redundant robots that combine the potential of asymmetric series-parallel actuated branches and multi-articulation pose a non-trivial challenge. To address the problem, this work proposes a novel optimization based controller that can accommodate various quadratic criteria to perform the torque distribution among dissimilar series and parallel actuators in order to maximize the motion efficiency. Three candidate criteria are composed and their performances are compared during periodic squat motions with a 3 degree of freedom series-parallel compliant articulated leg prototype. It is first shown that by minimizing a criterion that takes into account the actuator hardware specifications such as torque constant and transmission ratio, the gravity-driven phases can be lengthened. Thereby, this particular criterion results in slightly better performance than when adopting a strategy that maximizes the torque allocation to the higher efficiency actuators. Furthermore, valuable insights such as that the efficacy of maximum utilization of the highly-efficient parallel actuation branches decreases progressively at high frequencies were observed.

A quadcopter is an under-actuated system with only four control inputs for six degrees of freedom, and yet the human control of a quadcopter is simple enough to be learned with some practice. In this work, we consider the problem of human control of a multiple quadcopters system to transport a cable-suspended payload. The coupled dynamics of the system, due to the inherent physical constraints, is used to develop a leader-follower architecture where the leader quadcopter is controlled directly by a human operator and the followers are controlled with the proposed Payload Attitude Controller and Cable Attitude Controller. Experiments, where a human operator flew a two quadcopters system to transport a cable-suspended payload, were conducted to study the performance of proposed controller. The results demonstrated successful implementation of human control in these systems. This work presents the possibility of enabling manual control for on-the-go maneuvering of the quadcopter-payload system which motivates aerial transportation in the unknown environments.

Simplified models are useful to increase the computational efficiency of a motion planning algorithm, but their lack of accuracy have to be managed. We propose two feasibility constraints to be included in a Single Rigid Body Dynamics-based trajectory optimizer in order to obtain robust motions in challenging terrain. The first one finds an approximate relationship between joint-torque limits and admissible contact forces, without requiring the joint positions. The second one proposes a leg model to prevent leg collision with the environment. Such constraints have been included in a simplified nonlinear non-convex trajectory optimization problem. We demonstrate the feasibility of the resulting motion plans both in simulation and on the Hydraulically actuated Quadruped (HyQ) robot, considering experiments on an irregular terrain.

The role of various types of robot assistance in post-stroke gait rehabilitation has gained much attention in recent years. Furthermore, there is increased popularity to use more than one rehabilitation method in order to utilize the different advantages of each. Naturally, this results in the need to study how the different robot-assisted interventions affect the various underlying sensorimotor mechanisms involved in rehabilitation. To answer this important question, this paper combines a virtual reality experience with a unique robotic rehabilitation device, the Variable Stiffness Treadmill (VST), as a way of understanding interactions across different sensorimotor mechanisms involved in gait. The VST changes the walking surface stiffness in order to simulate real-world compliant surfaces while seamlessly interacting with a virtual environment. Through the manipulated visual and proprioceptive feedback, this paper focuses on the muscle activation patterns before, during, and after surface changes that are both visually informed and uninformed. The results show that there are predictable and repeatable muscle activation patterns both before and after surface stiffness changes, and these patterns are affected by the perceived visual and proprioceptive feedback. The interaction of feedback mechanisms and their effect on evoked muscular activation can be used in future robot-assisted gait therapies, where the intended muscle responses are informed by deterministic models and are tailored to a specific patient's needs.

The design of reliable path-following controllers is a key ingredient for successful deployment of self-driving vehicles. This controller-design problem is especially challenging for a general 2-trailer with a car-like tractor due to the vehicle's structurally unstable joint-angle kinematics in backward motion and the car-like tractor's curvature limitations which can cause the vehicle segments to fold and enter a jackknife state. Furthermore, advanced sensors with a limited field of view have been proposed to solve the joint-angle estimation problem online, which introduce additional restrictions on which vehicle states that can be reliably estimated. To incorporate these restrictions at the level of control, a model predictive path-following controller is proposed. By taking the vehicle's physical and sensing limitations into account, it is shown in real-world experiments that the performance of the proposed path-following controller in terms of suppressing disturbances and recovering from non-trivial initial states is significantly improved compared to a previously proposed solution where the constraints have been neglected.

We present a behaviour-based reinforcement learning approach, inspired by Brook's subsumption architecture, in which simple fully connected networks are trained as reactive behaviours. Our working assumption is that a pick and place robotic task can be simplified by leveraging domain knowledge of a robotics developer to decompose and train reactive behaviours; namely, approach, grasp, and retract. Then the robot autonomously learns how to combine reactive behaviours via an Actor-Critic architecture. We use an Actor-Critic policy to determine the activation and inhibition mechanisms of the reactive behaviours in a particular temporal sequence. We validate our approach in a simulated robot environment where the task is about picking a block and taking it to a target position while orienting the gripper from a top grasp. The latter represents an extra degree-of-freedom of which current end-to-end reinforcement learning approaches fail to generalise. Our findings suggest that robotic learning can be more effective if each behaviour is learnt in isolation and then combined them to accomplish the task. That is, our approach learns the pick and place task in 8,000 episodes, which represents a drastic reduction in the number of training episodes required by an end-to-end approach ( 95,000 episodes) and existing state-of-the-art algorithms.

A novel scheme for an "upgrade" of a linear control algorithm to a non-linear one is developed based on the concepts of a generalized homogeneity and an implicit homogeneous feedback design. Some tuning rules for a guaranteed improvement of a regulation quality are proposed. Theoretical results are confirmed by real experiments with the quadrotor QDrone of Quanser™.

This paper introduces the Omnidirectional Tractable Three Module Robot for traversing inside complex pipe networks. The robot consists of three omnidirectional modules fixed 120° apart circumferentially which can rotate about their own axis allowing holonomic motion of the robot. The holonomic motion enables the robot to overcome motion singularity when negotiating T-junctions and further allows the robot to arrive in a preferred orientation while taking turns inside a pipe. We have developed a closed-form kinematic model for the robot in the paper and propose the `Motion Singularity Region' that the robot needs to avoid while negotiating T-junction. The design and motion capabilities of the robot are demonstrated both by conducting simulations in MSC ADAMS on a simplified lumped-model of the robot and with experiments on its physical embodiment.