ICRA 2021 论文列表

The collaboration between humans and robots in an object search task requires the achievement of shared plans obtained from communicating and negotiating. In this work, we assume that the robot computes, as a first step, a multi-agent plan for both itself and the human. Then, both plans are submitted to human scrutiny, who either agrees or modifies it forcing the robot to adapt its own restrictions or preferences. This process is repeated along the search task as many times as required by the human. Our planner is based on a decentralized variant of Monte Carlo Tree Search (MCTS), with one robot and one human as agents. Moreover, our algorithm allows the robot and the human to optimize their own actions by maintaining a probability distribution over the plans in a joint-action space. The method allows an objective function definition over action sequences, it assumes intermittent communication, it is anytime and suitable for on-line replanning. To test it, we have developed a human-robot communication mobile phone interface. Validation is provided by real-life search experiments of a Parcheesi token in an urban space, including also an acceptability study.

Physical Human-Robot-Interaction (pHRI) is beneficial for communication in social interaction or to perform collaborative tasks but is also crucial for safety. While robotic devices embed sensors for this sole purpose, their design often is the results of a trade-off between technical capabilities and rarely considers human factors. We propose a novel approach to design and fabricate compliant Human-like artificial skin sensors for robots, with similar mechanical properties as human skin and capable of precisely detecting touch. Our artificial skin relies on the use of different silicone elastomers to replicate the human skin layers and comprises an embedded electrode matrix to perform mutual capacitance sensing. We present the sensor and describe its fabrication process which is scalable, low-cost and ensures flexibility, compliance and robustness. We introduce Muca, an open-source sensing development board and then evaluate the performance of the sensor.

Grasp planning and most specifically the grasp space exploration is still an open issue in robotics. This article presents an efficient procedure for exploring the grasp space of a multifingered adaptive gripper for generating reliable grasps given a known object pose. This procedure relies on a limited dataset of manually specified expert grasps, and use a mixed analytic and data-driven approach based on the use of a grasp quality metric and variational autoencoders. The performances of this method are assessed by generating grasps in simulation for three different objects. On this grasp planning task, this method reaches a grasp success rate of 99.91% on 7000 trials.

Energetically-passive robot exoskeletons, mimicking the function of the bicycle, could enable humans to reach previously unprecedented mobility. However, energetically-passive robot exoskeletons require a sophisticated mechanism to enable the human to supply energy, similar to what is enabled by the variable gear transmission mechanism of the bicycle. In this work, we present a new type of human-driven compliant transmission mechanism that could enable humans to supply energy when the leg is in the air, store the supplied energy, and release the stored energy when the leg is on the ground, in order to amplify the leg force and power. The compliant transmission mechanism presented in this paper is the first prototype and key component of a future human-driven artificial limb that aims to augment human mobility without using external energy.

This paper presents an experimental study that investigated how humans interact with viscous, damping-defined mechanical environments and quantified the lower bounds of robotic damping that they can stably interact with. Human subjects performed posture maintenance tasks for different arm postures while holding a robotic arm manipulator simulating unstable (negative) damping-defined environments and applying rapid perturbations to disturb the arm posture and challenge arm stability. The results of this study demonstrated that the lower bound of robotic damping for stable physical human-robot interaction was more than twice as low in the anterior-posterior (AP) direction than the medial-lateral (ML) direction, with lower limits of -50.3 Ns/m and -21.6 Ns/m in the AP and ML directions, respectively. The results further showed that the human arm is less capable of adjusting to the unstable environments when it is close to the body and laterally displaced for the AP and ML directions, respectively. Secondary analysis on the kinematic response in the phase space also demonstrated that arm stability in the unstable environments can be more easily achieved in the AP than ML direction. The outcomes of this study can be used to design less conservative robotic impedance or admittance controllers that utilize a wider range of robotic damping up to a certain extent of negative damping but do not compromise coupled stability of the human-robot system, which could improve the overall performance in physical human-robot interaction by achieving more agile operations and reducing user effort.

In this paper, HueCode, a meta-marker that robustly and simultaneously exposes the relative pose between a marker and a camera along with additional information, is proposed. It occupies the area of a single marker by overlaying multiple types of markers in different colored layers. Using perspective information from the first (most recognizable) type of element marker, the second or higher marker can be recognized with a better success rate. An experiment using a HueCode made from an ArUco marker and a QR code showed that the QR code within the HueCode is recognizable at an elevation angle of up to 15°, compared to 25° for a normal QR code. In addition, the versatility of HueCode is demonstrated using two robotic applications. The first is a lightweight estimation of absolute 6-DoF poses for mobile robots in a GNSS-denied environment, and the other is object annotation in two-dimensional image or three-dimensional space without any prior knowledge. The former realized pose estimation with a position error of less than 0.05 m, and the latter enabled annotation of a mirror and a transparent object, which are difficult for other sensors and machine learning to recognize.

Robot teleoperation is a reliable way to perform a variety of tasks with complex robotic systems. However, the remote control of active telepresence cameras on the robot for improved telepresence adds an additional degree of complexity while teleoperating and can thus affect the operator’s performance during tele-manipulation. Our previous user study investigates the general human performance and preference when using various wearable cameras. In this paper, we further investigate how humans respond to the usage of telepresence cameras in terms of motion behavior. The findings from our human motion analysis inform several desired designs for robot teleoperation interfaces and assistive autonomy.

Robots deployed in airports, malls, and stadiums today require expert oversight, pre-provided maps, and infrastructure. These systems are engineered to their specific deployment spaces and rely heavily on geometric maps for motion planning. In this work we consider a robot with only local-sensing and present a navigation strategy that uses signs and homotopy classes to fuel planning. We prove that in the worst-case, this strategy still maintains the probabilistic completeness that sampling based motion planners provide. Furthermore, we experimentally show that this exploration strategy results in 100% goal completion in real airport floor plans, and demonstrate how the robot’s sensing capabilities affects efficiency. We also show the effect of the environment variables, such as total number of obstacles, and density of obstacles, on navigation.

Despite their soft and compliant bodies, most of today’s soft robots have limitations when it comes to elongation or extension of their main structure. In contrast to this, a new type of soft robot called the eversion robot can grow longitudinally, exploiting the principle of eversion. Eversion robots can squeeze through narrow openings, giving the possibility to access places that are inaccessible by conventional robots. The main drawback of these types of robots is their limited bending capability due to the tendency to move along a straight line. In this paper, we propose a novel way to fuse bending actuation with the robot’s structure. We devise an eversion robot whose body forms both the central chamber that acts as the backbone as well as the actuators that cause bending and manoeuvre the manipulator. The proposed technique shows a significantly improved bending capability compared to externally attaching actuators to an eversion robot showing a 133% improvement in bending angle. Due to the increased manoeuvrability, the proposed solution is a step towards the employment of eversion robots in remote and difficult-to-access environments.

Line segment features become popular in SLAM community. Usually, line-based SLAM systems utilize local appearance descriptors for line segment tracking. However, traditional descriptor-based line segment tracking algorithms suffer from the problem that accuracy and speed cannot be possessed simultaneously, which affects the performance of line-based SLAM systems negatively. We propose a novel line segment tracking method with an IMU-KLT line segment prediction and a convex geometric distance minimization to boost line segment tracking performance in both accuracy and speed. Particularly, the proposed convex geometric distance minimization uses a ℓ1-norm model to minimize geometric constraints between predicted line segments and extracted line segments efficiently. Furthermore, the line segment tracking is embedded into a VIO system and we adapt it to obtain more reliable point tracking. Experimental results on public datasets show that the proposed line segment tracking method achieves much higher accuracy and much less time cost than state-of-the-art level, where not only the number of correct matches increases but also the inlier ratios are increased by at least 35.1% along with a 3 times faster speed. Besides, the VIO system combining the proposed line segment tracking is improved in terms of accuracy.

Safe and high-speed navigation is a key enabling capability for real world deployment of robotic systems. A significant limitation of existing approaches is the computational bottleneck associated with explicit mapping and the limited field of view (FOV) of existing sensor technologies. In this paper, we study algorithmic approaches that allow the robot to predict spaces extending beyond the sensor horizon for robust planning at high speeds. We accomplish this using a generative neural network trained from real-world data without requiring human annotated labels. Further, we extend our existing control algorithms to support leveraging the predicted spaces to improve collision-free planning and navigation at high speeds. Our experiments are conducted on a physical robot based on the MIT race car using an RGBD sensor where were able to demonstrate improved performance at 4 m/s compared to a controller not operating on predicted regions of the map.

In this paper, we propose a new planning scheme for high-speed flight in an unknown environment while taking into account drag forces. Drag forces become non-negligible at high speeds and may lead to unfeasible trajectories. The method leverages a new Mixed-Integer Quadratic Program/Model Predictive Control formulation that allows to easily account for drag forces. This formulation makes use of a state-of-the-art Safe Corridors generator to guarantee safety. It uses state-of-the-art mapping algorithms and solvers to achieve a higher computational efficiency than similar state-of-the-art methods. To the best of our knowledge, our method is the first high-speed planner that generates safe trajectories while accounting for drag. The proposed method is tested in simulation and compared to similar state-of-the-art methods for planning in unknown environments in terms of quality and computation time.

Model Predictive Control (MPC) promises to endow robots with enough reactivity to perform complex tasks in dynamic environments by frequently updating their motion plan based on measurements. Despite its appeal, it has seldom been deployed on real machines because of scaling constraints. This paper presents the first hardware implementation of closed-loop nonlinear MPC on a 7-DoF torque-controlled robot. Our controller leverages a state-of-the art optimal control solver, namely Differential Dynamic Programming (DDP), in order to replan state and control trajectories at real-time rates (1kHz). In addition to this experimental proof of concept, an exhaustive performance analysis shows that our controller outperforms open-loop MPC on a rapid cyclic end-effector task. We also exhibit the importance of a sufficient preview horizon and full robot dynamics through comparisons with inverse dynamics and kinematic optimization.

Solving robotic navigation tasks via reinforcement learning (RL) is challenging due to their sparse reward and long decision horizon nature. However, in many navigation tasks, high-level (HL) task representations, like a rough floor plan, are available. Previous work has demonstrated efficient learning by hierarchal approaches consisting of path planning in the HL representation and using sub-goals derived from the plan to guide the RL policy in the source task. However, these approaches usually neglect the complex dynamics and sub-optimal sub-goal-reaching capabilities of the robot during planning. This work overcomes these limitations by proposing a novel hierarchical framework that utilizes a trainable planning policy for the HL representation. Thereby robot capabilities and environment conditions can be learned utilizing collected rollout data. We specifically introduce a planning policy based on value iteration with a learned transition model (VI-RL). In simulated robotic navigation tasks, VI-RL results in consistent strong improvement over vanilla RL, is on par with vanilla hierarchal RL on single layouts but more broadly applicable to multiple layouts, and is on par with trainable HL path planning baselines except for a parking task with difficult non-holonomic dynamics where it shows marked improvements.

This paper deals with the traffic management of multiple Automated Guided Vehicles (AGVs) in an automatic factory or warehouse. We propose innovative methods, evolved from the studies previously conducted in [1], to coordinate a fleet of AGVs in an industrial environment, and we describe the methodologies developed to build a complete traffic manager software. The software is based on a multi-layer control architecture: a higher-level layer useful to model the traffic of vehicles among the different areas of the warehouse, a middle layer which acts as a bridge between the traffic model and the path planner, and a lower-level layer which represents the roadmap itself and, hence, defines the optimal path to be followed by each vehicle. Finally, the AGVs movement coordination is managed separately through a centralized control in order to avoid conflicts and deadlocks.The aim is to make theoretical methods applicable to a real environment facing the usual problems related to the industrial applications, which have been overlooked in [1]. Indeed, the roadmap is usually constrained by the plant layout, especially in medium size factories, and paths can not be arbitrarily defined. Hence, the aim is to realize a reliable and robust software able to manage real scenarios, allowing the traffic management of multiple AGVs. The proposed method aims at flexibility by considering a coordination strategy not based on assumptions and ad-hoc rules.

We present a visually grounded hierarchical planning algorithm for long-horizon manipulation tasks. Our algorithm offers a joint framework of neuro-symbolic task planning and low-level motion generation conditioned on the specified goal. At the core of our approach is a two-level scene graph representation, namely geometric scene graph and symbolic scene graph. This hierarchical representation serves as a structured, object-centric abstraction of manipulation scenes. Our model uses graph neural networks to process these scene graphs for predicting high-level task plans and low-level motions. We demonstrate that our method scales to long-horizon tasks and generalizes well to novel task goals. We validate our method in a kitchen storage task in both physical simulation and the real world. Experiments show that our method achieves over 70% success rate and nearly 90% of subgoal completion rate on the real robot while being four orders of magnitude faster in computation time compared to standard search-based task-and-motion planner. 1

We propose a hierarchical representation of objects, where the representation of each object is allowed to change based on the quality of accumulated measurements. We initially estimate each object as a 2D bounding box or a 3D point, encoding only the geometric properties that can be well-constrained using limited viewpoints. With additional measurements, we allow each object to become a higher dimensional 3D volumetric model for improved reconstruction accuracy and collision-testing. Our Hierarchical Object Map Estimation (HOME) is robust to deficiencies in viewpoints and allows planning safe and efficient trajectories around object obstacles using a monocular camera. We demonstrate the advantages of our approach on a real-world TUM dataset and during visual-inertial navigation of a quad-rotor in simulation.

Systems of multiple low-cost, underactuated floats combined with fully actuated surface vessels can improve the scalability and cost-effectiveness of autonomous systems for marine science and environmental monitoring. Here, we consider a coordination problem where surface vessels must drop off floats at locations such that they are likely to drift to observe given points of interest, and later must pick up the floats for redeployment. We define the Multi-Vessel Multi-Float (MVMF) problem and present a hierarchical solution based on the Dec-MCTS algorithm. Our solution defines customised sampling, rollout, and action generation algorithms to accommodate the problem’s large search space and provide computational performance sufficient for practical application. We report analytical and simulation results that demonstrate the computational efficiency of our method and validate its behaviour in practical problems. These results immediately enable field experiments to progress the development of this exciting concept in multi-robot marine systems.

Although reinforcement learning (RL) has achieved great success in robotic manipulation skills learning, it is still challenging for long-horizon tasks. Combining RL with demonstrations is an effective solution. In this paper, we propose a novel hierarchical learning from demonstrations method for long-horizon tasks, which leverages (i) object-centered segmentation of demonstrations to automatically segment the teaching trajectories into episodes. (ii) a bi-level hierarchical imitation learning method with a parallel training mechanism to train the two-level policies simultaneously. Experimental results on three challenging long-horizon tasks with sparse rewards show that our proposed method significantly outperforms state-of-art approaches in terms of both sample-efficiency and success rate. Moreover, our method is the only one which achieves satisfactory performance in tasks of multi-object stack and multi-object push&stack.