ICRA 2020 论文列表

We present a distributed Bayesian algorithm for robot swarms to classify a spatially distributed feature of an environment. This type of "go/no-go" decision appears in applications where a group of robots must collectively choose whether to take action, such as determining if a farm field should be treated for pests. Previous bio-inspired approaches to decentralized decision-making in robotics lack a statistical foundation, while decentralized Bayesian algorithms typically require a strongly connected network of robots. In contrast, our algorithm allows simple, sparsely distributed robots to quickly reach accurate decisions about a binary feature of their environment. We investigate the speed vs. accuracy tradeoff in decision-making by varying the algorithm's parameters. We show that making fewer, less-correlated observations can improve decision-making accuracy, and that a well-chosen combination of prior and decision threshold allows for fast decisions with a small accuracy cost. Both speed and accuracy also improved with the addition of bio-inspired positive feedback. This algorithm is also adaptable to the difficulty of the environment. Compared to a fixed-time benchmark algorithm with accuracy guarantees, our Bayesian approach resulted in equally accurate decisions, while adapting its decision time to the difficulty of the environment.

Many species have evolved advanced non-visual perception while artificial systems fall behind. Radar and ultrasound complement camera-based vision but they are often too costly and complex to set up for very limited information gain. In nature, sound is used effectively by bats, dolphins, whales, and humans for navigation and communication. However, it is unclear how to best harness sound for machine perception.Inspired by bats' echolocation mechanism, we design a low- cost BatVision system that is capable of seeing the 3D spatial layout of space ahead by just listening with two ears. Our system emits short chirps from a speaker and records returning echoes through microphones in an artificial human pinnae pair. During training, we additionally use a stereo camera to capture color images for calculating scene depths. We train a model to predict depth maps and even grayscale images from the sound alone. During testing, our trained BatVision provides surprisingly good predictions of 2D visual scenes from two 1D audio signals. Such a sound to vision system would benefit robot navigation and machine vision, especially in low-light or no-light conditions. Our code and data are publicly available.

In this work, we demonstrate an instrumented wheel concept which utilizes a 2D pressure grid, an electrochemical impedance spectroscopy (EIS) sensor and machine learning (ML) to extract meaningful metrics from the interaction between the wheel and surface terrain. These include continuous slip/skid estimation, balance, and sharpness for engineering applications. Estimates of surface hydration, texture, terrain patterns, and regolith physical properties such as cohesion and angle of internal friction are additionally calculated for science applications. Traditional systems rely on post-processing of visual images and vehicle telemetry to estimate these metrics. Through in-situ sensing, these metrics can be calculated in near real time and made available to onboard science and engineering autonomy applications. This work aims to provide a deployable system for future planetary exploration missions to increase science and engineering capabilities through increased knowledge of the terrain.

We study a novel class of motion planning problems, inspired by emerging low-energy robotic vehicles, such as insect-size flyers, chip-size satellites, and high-endurance autonomous blimps, for which the energy consumed by computing hardware during planning a path can be as large as the energy consumed by actuation hardware during the execution of the same path. We propose a new algorithm, called Compute Energy Included Motion Planning (CEIMP). CEIMP operates similarly to any other anytime planning algorithm, except it stops when it estimates further computing will require more computing energy than potential savings in actuation energy. We show that CEIMP has the same asymptotic computational complexity as existing sampling-based motion planning algorithms, such as PRM*. We also show that CEIMP outperforms the average baseline of using maximum computing resources in realistic computational experiments involving 10 floor plans from MIT buildings. In one representative experiment, CEIMP outperforms the average baseline 90.6% of the time when energy to compute one more second is equal to the energy to move one more meter, and 99.7% of the time when energy to compute one more second is equal to or greater than the energy to move 3 more meters.

This study deals with the balance of humanoid or multi-legged robots in a multi-contact setting where a chosen subset of contacts is undergoing desired sliding-task motions. One method to keep balance is to hold the center-of-mass (CoM) within an admissible convex area. This area is calculated based on the contact positions and forces. We introduce a methodology to compute this CoM support area (CSA) for multiple fixed and intentionally sliding contacts. To select the most appropriate CoM position within CSA, we account for (i) constraints of multiple fixed and sliding contacts, (ii) desired wrench distribution for contacts, and (iii) desired CoM position (eventually dictated by other tasks). These are formulated as a quadratic programming (QP) optimization problems. We illustrate our approach with pushing against a wall and wiping, and conducted experiments using the HRP-4 humanoid robot.

Series elastic actuators with passive compliance have been gaining increasing popularity in force-controlled robotic manipulators. One of the reasons is the actuator's ability to infer the applied torque by measuring the deflection of the elastic element as opposed to directly with dedicated torque sensors. Proper deflection control is pinnacle to achieve a desired output torque and, therefore, small deviances in positional measurements or a nonlinear deformation can have adverse effects on performance. In applications with larger torque requirements, the actuators typically use gear reductions which inherently result in mechanical backlash. This combined with the nonlinear behaviour of the elastic element and unmodelled dynamics, can severely compromise force fidelity.This paper proposes a backlash compensating active disturbance rejection controller (ADRC) for multi-input series elastic actuators. In addition to proper deflection control, a multiinput active disturbance rejection controller is derived and implemented experimentally to mitigate any unmodelled nonlinearities or perturbations to the plant model. The controller is experimentally validated on a hybrid motor-brake-clutch series elastic actuator and the controller performance is compared against traditional error-based controllers. It is shown that the backlash compensated ADRC outperforms classical PID and ADRC methods and is a viable solution to positional measurement error in elastic actuators.

This paper introduces a new approach for victim localization in avalanches that will be exploited by UAVs using the ARVA sensor. We show that the nominal ARVA measurement can be linearly related to a quantity that is sufficient to reconstruct the victim position. We explicitly deal with a robust scenario in which the measurement is actually perturbed by a noise that grows with the distance to the victim and we propose an adaptive control scheme made of a least-square identifier and a trajectory generator whose role is both to guarantee the persistence of excitation for the identifier and to steer the ARVA receiver towards the victim. We show that the system succeeds in localizing the victim in a domain where the ARVA output is sufficiently informative and illustrate its performance in simulation. This new approach could significantly reduce the searching time by providing an exploitable estimate before having reached the victim. The work is framed within the EU project AirBorne whose goals is to develop at TRL8 a drone for quick localization of victims in avalanche scenarios.

A fundamental challenge in retinal surgery is safely navigating a surgical tool to a desired goal position on the retinal surface while avoiding damage to surrounding tissues, a procedure that typically requires tens-of-microns accuracy. In practice, the surgeon relies on depth-estimation skills to localize the tool-tip with respect to the retina in order to perform the tool-navigation task, which can be prone to human error. To alleviate such uncertainty, prior work has introduced ways to assist the surgeon by estimating the tooltip distance to the retina and providing haptic or auditory feedback. However, automating the tool-navigation task itself remains unsolved and largely unexplored. Such a capability, if reliably automated, could serve as a building block to streamline complex procedures and reduce the chance for tissue damage. Towards this end, we propose to automate the tool-navigation task by learning to mimic expert demonstrations of the task. Specifically, a deep network is trained to imitate expert trajectories toward various locations on the retina based on recorded visual servoing to a given goal specified by the user. The proposed autonomous navigation system is evaluated in simulation and in physical experiments using a silicone eye phantom. We show that the network can reliably navigate a needle surgical tool to various desired locations within 137 μm accuracy in physical experiments and 94 μm in simulation on average, and generalizes well to unseen situations such as in the presence of auxiliary surgical tools, variable eye backgrounds, and brightness conditions.

Autonomous tool construction is a significant but challenging task in robotics. This task can be interpreted as when given a reference tool, selecting some available candidate parts to reconstruct it. Most of the existing works perform tool construction in the form of action part and grasp part, which is only a specific construction pattern and limits its application to some extent. In general scenarios, a tool can be constructed in various patterns with different part pairs. Therefore, whether a part pair is most suitable for constructing the tool depends not only on itself, but on other parts in the same scene. To solve this problem, we construct a Gated Graph Neural Network (GGNN) to model the relations between all part pairs, so that we can select the candidate parts in consideration of the global information. Afterwards, we embed the constructed GGNN into a RCNN-like structure to finally accomplish tool construction. The whole model will be named Tool Construction Graph RCNN (TC-GRCNN). In addition, we develop a mechanism that can generate large-scale training and testing data in simulation environments, by which we can save the time of data collection and annotation. Finally, the proposed model is deployed on the physical robot. The experiment results show that TC-GRCNN can perform well in the general scenarios of tool construction.

In Minimally Invasive Surgery (MIS), tissue scanning with imaging probes is required for subsurface visualisation to characterise the state of the tissue. However, scanning of large tissue surfaces in the presence of motion is a challenging task for the surgeon. Recently, robot-assisted local tissue scanning has been investigated for motion stabilisation of imaging probes to facilitate the capturing of good quality images and reduce the surgeon's cognitive load. Nonetheless, these approaches require the tissue surface to be static or translating with periodic motion. To eliminate these assumptions, we propose a visual servoing framework for autonomous tissue scanning, able to deal with free-form tissue motion. The 3D structure of the surgical scene is recovered, and a feature-based method is proposed to estimate the motion of the tissue in real-time. The desired scanning trajectory is manually defined on a reference frame and continuously updated using projective geometry to follow the tissue motion and control the movement of the robotic arm. The advantage of the proposed method is that it does not require the learning of the tissue motion prior to scanning and can deal with free-form motion. We deployed this framework on the da Vinci®surgical robot using the da Vinci Research Kit (dVRK) for Ultrasound tissue scanning. Our framework can be easily extended to other probe-based imaging modalities.

This paper focuses on real-time occupancy mapping and collision checking onboard an autonomous robot navigating in an unknown environment. We propose a new map representation, in which occupied and free space are separated by the decision boundary of a kernel perceptron classifier. We develop an online training algorithm that maintains a very sparse set of support vectors to represent obstacle boundaries in configuration space. We also derive conditions that allow complete (without sampling) collision-checking for piecewise-linear and piecewise-polynomial robot trajectories. We demonstrate the effectiveness of our mapping and collision checking algorithms for autonomous navigation of an Ackermann-drive robot in unknown environments.

The ability to autonomously modify their environment dramatically increases the capability of robots to operate in unstructured environments. We develop a specialized construction algorithm and robotic system that can autonomously build motion support structures with previously unseen objects. The approach is based on our prior work on adaptive ramp building algorithms, but it eliminates the assumption of having specialized building materials that simplify manipulation and planning for stability. Utilizing irregularly shaped stones makes the problem significantly more challenging since the outcome of individual placements is sensitive to details of contact geometry and friction, which are difficult to observe. To reuse the same high-level algorithm, we develop a new physics-based planner that explicitly considers the uncertainty produced by incomplete in-situ sensing and imprecision during pickup and placement. We demonstrate the approach on a robotic system that uses a newly developed gripper to reliably pick up stones with minimal additional sensors or complex grasp planning. The resulting system can build structures with more than 70 stones, which in turn provide traversable paths to previously inaccessible locations.

Robot simulators provide an easy way for evaluation of new concepts and algorithms in a simulated physical environment reducing development time and cost. Therefore it is convenient to have a tool that quickly creates a 3D landscape from an arbitrary 2D image or 2D laser range finder data. This paper presents a new tool that automatically constructs such landscapes for Gazebo simulator. The tool converts a grayscale image into a 3D Collada format model, which could be directly imported into Gazebo. We run three different simultaneous localization and mapping (SLAM) algorithms within three varying complexity environments that were constructed with our tool. A real-time factor (RTF) was used as an efficiency benchmark. Successfully completed SLAM missions with acceptable RTF levels demonstrated the efficiency of the tool. The source code is available for free academic use.

In this paper, we propose a method for generating undulatory gaits for snake robots. Instead of starting from a pre-defined movement pattern such as a serpenoid curve, we use a Model Predictive Control (MPC) approach to automatically generate effective locomotion gaits via trajectory optimization. An important advantage of this approach is that the resulting gaits are automatically adapted to the environment that is being modeled as part of the snake dynamics. To illustrate this, we use a novel model for anisotropic dry friction, along with existing models for viscous friction and fluid dynamic effects such as drag and added mass. For each of these models, gaits generated without any change in the method or its parameters are as efficient as Pareto-optimal serpenoid gaits tuned individually for each environment. Furthermore, the proposed method can also produce more complex or irregular gaits, e.g. for obstacle avoidance or executing sharp turns.

This paper investigates a novel calibration process of devices with different modalities, which is a critical step of computer vision applications. We propose a fully automatic extrinsic calibration of a LiDAR-camera system. Our approach applies sphere as their surfaces and contours can be accurately detected on point clouds and camera images, respectively. Experiments on synthetic and real data exhibit that our automatic algorithm is fast and robust and it yields accurate camera and LiDAR extrinsic parameters.

Automatic surgical gesture recognition is fundamental for improving intelligence in robot-assisted surgery, such as conducting complicated tasks of surgery surveillance and skill evaluation. However, current methods treat each frame individually and produce the outcomes without effective consideration on future information. In this paper, we propose a framework based on reinforcement learning and tree search for joint surgical gesture segmentation and classification. An agent is trained to segment and classify the surgical video in a human-like manner whose direct decisions are re-considered by tree search appropriately. Our proposed tree search algorithm unites the outputs from two designed neural networks, i.e., policy and value network. With the integration of complementary information from distinct models, our framework is able to achieve the better performance than baseline methods using either of the neural networks. For an overall evaluation, our developed approach consistently outperforms the existing methods on the suturing task of JIGSAWS dataset in terms of accuracy, edit score and F1 score. Our study highlights the utilization of tree search to refine actions in reinforcement learning framework for surgical robotic applications.

Current challenges in automated robotic breast ultrasound (US) acquisitions include keeping acoustic coupling between the breast and the US probe, minimizing tissue deformations and safety. In this paper, we present how an autonomous 3D breast US acquisition can be performed utilizing a 7DOF robot equipped with a linear US transducer. Robotic 3D breast US acquisitions would increase the diagnostic value of the modality since they allow patient specific scans and have a high reproducibility, accuracy and efficiency. Additionally, 3D US acquisitions allow more flexibility in examining the breast and simplify registration with preoperative images like MRI. To overcome the current challenges, the robot follows a reference- based trajectory adjusted by a visual servoing algorithm. The reference trajectory is a patient specific trajectory coming from e.g. an MRI. The visual servoing algorithm commands in-plane rotations and corrects the probe contact based on confidence maps. A safety aware, intrinsically passive framework is utilised to actuate the robot. The approach is illustrated with experiments on a phantom, which show that the robot only needs minor pre-procedural information to consistently image the phantom while relying mainly on US feedback.

Exosuits can reduce metabolic demand and improve gait. Controllers explicitly derived from biological mechanisms that reflect the user's joint or muscle dynamics should in theory allow for individualized assistance and enable adaptation to changing gait. With the goal of developing an exosuit control strategy based on muscle power, we present an approach for estimating, at real time rates, when the soleus muscle begins to generate positive power. A low-profile ultrasound system recorded B-mode images of the soleus in walking individuals. An automated routine using optical flow segmented the data to a normalized gait cycle and estimated the onset of concentric contraction at real-time rates (~130Hz). Segmentation error was within 1% of the gait cycle compared to using ground reaction forces. Estimation of onset of concentric contraction had a high correlation (R2=0.92) and an RMSE of 2.6% gait cycle relative to manual estimation. We demonstrated the ability to estimate the onset of concentric contraction during fixed speed walking in healthy individuals that ranged from 39.3% to 45.8% of the gait cycle and feasibility in two persons post-stroke walking at comfortable walking speed. We also showed the ability to measure a shift in onset timing to 7% earlier when the biological system adapts from level to incline walking. Finally, we provided an initial evaluation for how the onset of concentric contraction might be used to inform exosuit control in level and incline walking.

In this paper, an automated tracking system with head and tail recognition for time-lapse observation of free-moving C. elegans is presented. In microscale field, active C. elegans can move out of the view easily without an automated tracking system because of the narrow field of view and rapid speed of C. elegans. In our previous works, we constructed an automated platform with 3D freedom to track centroid region of the nematode successfully. However, tracking time was not long enough to support a full time-lapse observation. Our proposed system in this study integrate the detection method in horizontal plane with depth evaluation more tightly. Tracking time and response speed have been greatly improved. Besides, we make full use of curvature calculation to make the system recognize the head and tail of C. elegans and the recognition rate can be up to 95%. The results demonstrate that the system can fully achieve automated long-term tracking of a free-living nematode and will be a nice tool for C. elegans behavioral analysis.