IROS 2021 论文列表

A dynamic autonomy allocation framework automatically shifts how much control lies with the human versus the robotics autonomy, for example based on factors such as environmental safety or user preference. To investigate the question of which factors should drive dynamic autonomy allocation, we perform a human subject study to collect ground truth data that shifts between levels of autonomy during shared-control robot operation. Information streams from the human, the interaction between the human and the robot, and the environment are analyzed. Machine learning methods—both classical and deep learning—are trained on this data. An analysis of information streams from the human-robot team suggests features which capture the interaction between the human and the robotics autonomy are the most informative in predicting when to shift autonomy levels. Even the addition of data from the environment does little to improve upon this predictive power. The features learned by deep networks, in comparison to the hand-engineered features, prove variable in their ability to represent shift-relevant information. This work demonstrates the classification power of human-only and human-robot interaction information streams for use in the design of shared-control frameworks, and provides insights into the comparative utility of various data streams and methods to extract shift-relevant information from those data.

Collecting new experience is costly in many robotic tasks, so determining how to efficiently explore in a new environment to learn as much as possible in as few trials as possible is an important problem for robotics. In this paper, we propose a method for exploring for the purpose of learning a dynamics model. Our key idea is to minimize a score given by a discriminator network as an objective for a planner which chooses actions. This discriminator is optimized jointly with a prediction model and enables our active learning approach to sample sequences of observations and actions which result in predictions considered the least realistic by the discriminator. Comparable existing exploration methods cannot operate in many prediction-planning pipelines used in robotic learning without hardware modifications to standard robotics platforms in order to accommodate their large compute requirements, so the primary contribution of our adversarial exploration method is scalability. We demonstrate progressively increased performance of our adversarial exploration approach compared to leading model-based exploration strategies as compute is restricted in simulated environments. We further demonstrate the ability of our adversarial method to scale to a robotic manipulation prediction-planning pipeline where we improve sample efficiency and prediction performance for a domain transfer problem.

Machines tend to use powerful actuators and large gearboxes to bear large loads, which are inconvenient in terms of responsiveness as they affect the duration of operations. Thus, to compensate the force to grasp an object, we propose a clamping mechanism implementing the internally-balanced magnetic unit (IB Magnet) as a force amplifier, which is a mechanism able to switch attached and detached states of a permanent magnet with an external force considerably smaller than its original attractive force. To realize the bi-parting constitution of fingers, a new compensation method using conical coil springs was designed to provide both precision and miniaturization. Relative to the constitution with a single motored screw, the prototype gripper for proof of concept successfully amplified the grasping force at most to 292.2% assisted by the magnetic attraction, while keeping the increase in power consumption of a DC motor only by 11.8%, making the force-energy efficiency 2.6 times larger. Thus, it was verified that the proposed gripper enables the use of actuators and current supplies that require less power.

Drone audition, namely the hearing capability of a drone, is expected to compensate for the drawbacks of visual sensors in search-and-rescue missions. Current multi-rotor drones have limitations of flight duration and sound processing due to ego-noise generated by rotors and air-flow. Drone audition for a kiteplane, i.e., a fixed-wing drone that can fly slowly and stably, has not been investigated. This paper proposes "Alternating Drive-and-Glide Flight Navigation"(AltDGFNavi) of a kiteplane for sound source position estimation. AltDGFNavi consists of two functions: periodical switching rotor for driving and gliding to reduce ego-noise, and dynamic flight path generation to fly close to the target. AltDGFNavi was evaluated through numerical simulations, and the results of sound source position estimation demonstrated the effectiveness of AltDGFNavi.

Single image haze removal is crucial in computer vision. In open literatures, two kinds of dehazing strategies (prior-based and learning-based methods) have been developed. However, they have a trade-off between detail preservation and the image quality. Prior-based methods reconstruct the detail well but have lower image quality while learning-based methods achieve better recovered quality but lose the detail. In this paper, to mitigate this dilemma, a hierarchical architecture using the discrete wavelet transform (DWT) is proposed. It divides the dehazing problem into two parts: detail and background reconstruction. Based on investigating how haze affects the image in the wavelet domain, two networks for detail and background reconstruction are proposed. To avoid color distortion and the detail loss, the anti-vanish wavelet loss and the bound penalty are proposed. The multi-level wavelet component discriminator is proposed for further improvement. Experiments show that the proposed network can achieve superior performance in all metrics.

Autonomous Micro Aerial Vehicles (MAVs) have the potential to be employed for surveillance and monitoring tasks. By perching and staring on one or multiple locations aerial robots can save energy while concurrently increasing their overall mission time without actively flying. In this paper, we address the estimation, planning, and control problems for autonomous perching on inclined surfaces with small quadrotors using visual and inertial sensing. We focus on planning and executing dynamically feasible trajectories to navigate and perch to a desired target location with on board sensing and computation. Our planner also supports certain classes of nonlinear global constraints by leveraging an efficient algorithm that we have mathematically verified. The on board cameras and IMU are concurrently used for state estimation and to infer the relative robot/target localization. The proposed solution runs in real-time on board a limited computational unit. Experimental results validate the proposed approach by tackling aggressive perching maneuvers with flight envelopes that include large excursions from the hover position on inclined surfaces up to 90°, angular rates up to 600 deg/s, and accelerations up to 10 m/s2.

Autonomous systems often operate in environments where the behavior of multiple agents is coordinated by a shared global state. Reliable estimation of the global state is thus critical for successfully operating in a multi-agent setting. We introduce agent-aware state estimation—a framework for calculating indirect estimations of state given observations of the behavior of other agents in the environment. We also introduce transition-independent agent-aware state estimation—a tractable class of agent-aware state estimation—and show that it allows the speed of inference to scale linearly with the number of agents in the environment. As an example, we model traffic light classification in instances of complete loss of direct observation. By taking into account observations of vehicular behavior from multiple directions of traffic, our approach exhibits accuracy higher than that of existing traffic light-only HMM methods on a real-world autonomous vehicle data set under a variety of simulated occlusion scenarios.

Learning from human feedback using event-related electroencephalography (EEG) signals has attracted extensive attention recently owing to their intuitive communication ability by decoding user intentions. However, this approach requires users to perform specified tasks and their success or failure. In addition, the amount of attention needed for decision-making increases with the task difficulty, decreasing human feedback quality over time because of fatigue. Consequently, this can reduce the interaction quality and can even cause interaction breakdowns. To overcome these limitations and enable the interaction of robots with higher complexity tasks, we propose a closed-loop control system that learns affective responses to robot behaviors and provides natural feedback to optimize robot parameters for smoothing the next action. Experimental results demonstrate our affect-driven closed-loop control system yielded better affective outcomes and task performance than an open-loop system with correlated neuroscientific characteristics of EEG signals, thus enhancing the quality of human-robot interaction.

The beneficial aspects of fully-actuated multirotor UAVs, provided by nonparallel rotor configuration, are increasingly being recognized and utilized to great benefit in high-precision applications. Full six-degree-of-freedom force control, higher control bandwidth and improved disturbance rejection prove valuable. However, the cant angle will cause great multirotor dihedral effect and significantly affects blade flapping, which decreases the flight performance of nonparallel actuated UAVs. Therefore, this paper presents a novel aerodynamic model for fully-actuated hexrotor UAVs while considering the aerodynamic effects caused due to tilt angled propeller configurations. In the proposed aerodynamic model, the significance of multirotor dihedral effect, defined as an aerodynamic coefficient proportional to the relative linear velocity of the UAV, is modeled for nonparallel actuators. Additionally, the modeling for blade flapping effect for cant angled propellers is provided to accurately model the aerodynamics. Wind tunnel experiments were conducted to characterize the aerodynamic constants for multirotor dihedral effect, blade flapping effect and air friction. Experimental results are presented to validate the proposed aerodynamic model on a fully-actuated hexrotor UAV (Purdue’s Dexterous Hexrotor). Lastly, the multirotor dihedral effect and blade flapping effect at different cant angles and at different wind speeds are analyzed.

Since omnidirectional aerial vehicles can generate a six degrees of freedom wrench, they could be used for dexterous manipulation tasks without the need for an additional robotic arm. However, they suffer from a reduced efficiency and dynamics range due to the huge amount of energy lost in gravity compensation.In this work, we introduce an omnidirectional aerial manipulator suspended from a cable-driven parallel robot (CDPR) by a spring, combining the advantages of the CDPR large workspace with the high dynamics of aerial vehicles, while reducing energy consumption thanks to gravity compensation.A partitioned control scheme is implemented to regulate both systems separately. A preliminary control strategy is proposed for the CDPR motion that minimizes the total energy consumption. Experiments are carried out to assess the added value of the CDPR carrier.

Most autonomous vehicles (AVs) rely on LiDAR and RGB camera sensors for perception. Using these point cloud and image data, perception models based on deep neural nets (DNNs) have achieved state-of-the-art performance in 3D detection. The vulnerability of DNNs to adversarial attacks have been heavily investigated in the RGB image domain and more recently in the point cloud domain, but rarely in both domains simultaneously. Multi-modal perception systems used in AVs can be divided into two broad types: cascaded models which use each modality independently, and fusion models which learn from different modalities simultaneously. We propose a universal and physically realizable adversarial attack for each type, and study and contrast their respective vulnerabilities to attacks. We place a single adversarial object with specific shape and texture on top of a car with the objective of making this car evade detection. Evaluating on the popular KITTI benchmark, our adversarial object made the host vehicle escape detection by each model type more than 50% of the time. The dense RGB input contributed more to the success of the adversarial attacks on both cascaded and fusion models.

Behavioral cloning (BC) bears a high potential for safe and direct transfer of human skills to robots. However, demonstrations performed by human operators often contain noise or imperfect behaviors that can affect the efficiency of the imitator if left unchecked. In order to allow the imitators to effectively learn from imperfect demonstrations, we propose to employ the robust t-momentum optimization algorithm. This algorithm builds on the Student's t-distribution in order to deal with heavy-tailed data and reduce the effect of outlying observations. We extend the t-momentum algorithm to allow for an adaptive and automatic robustness and show empirically how the algorithm can be used to produce robust BC imitators against datasets with unknown heaviness. Indeed, the imitators trained with the t-momentum-based Adam optimizers displayed robustness to imperfect demonstrations on two different manipulation tasks with different robots and revealed the capability to take advantage of the additional data while reducing the adverse effect of non-optimal behaviors.

This paper presents an adaptive backstepping controller for the reference tracking of an alginate artificial cell. An adaptive controller was implemented to precisely manipulate a magnetic artificial cell actuated by rotating magnetic fields. The rolling motion of a small-scale robot in a fluidic environment is challenging, especially when the fluid imparts an unknown response at low Reynolds number. In order to compensate for this uncertainty, an unknown tuning parameter encapsulating these effects was added to the governing equations of motion. A controller with an update law was then designed to estimate the unknown parameter and force the artificial cell to produce the desired response. The stability of the proposed controller was established by a candidate Lyapunov function. Real-time experiments were conducted to demonstrate the effectiveness of the designed controller at guiding an artificial cell to an arbitrary target position. Alginate cells were guided through a maze using the controller and was later combined with wall constraints to allow multiple alginate cells to reach the same target location. This controller can be applied to both surface motion and swimming-based small-scale robots in future applications for micro-assembly and targeted drug delivery.

This study addresses the challenge of predicting the terrain traversability of off-road vehicles. When an off-road vehicle is operated on rough terrains or slopes of unconsolidated materials, it is crucial to accurately predict terrain traversability for efficient operations and to avoid critical mobility risks. However, the prediction of traversability is challenging, especially for the prediction of possibly risky terrains because for such terrains, the traverse data available is either limited or non-existent. To address this limitation, this study proposes an adaptive terrain traversability prediction method based on the multi-source transfer Gaussian process regression (MS-TGPR). The proposed method utilizes limited data available on low risk terrains of the target environment to enhance the prediction accuracy by leveraging past traverse experiences on multiple types of terrain surfaces. The effectiveness of the proposed method is demonstrated using a slip dataset of various terrain surfaces and geometries.

Autonomous vehicles usually consume a large amount of computational power for their operations, especially for the tasks of sensing and perception with artificial intelligence algorithms. Such a computation may not only cost a significant amount of energy but also cause performance issues when the onboard computational resources are limited. To address this issue, this paper proposes an adaptive optimization method to online allocate the onboard computational resources of an autonomous vehicle amongst multiple vehicular subsystems depending on the contexts of the situations that the vehicle is facing. Different autonomous driving scenarios were designed to validate the proposed approach and the results showed that it could help improve the overall performance and energy consumption of autonomous vehicles compared to existing computational arrangement.

This study proposes an adaptive data-driven hy-perparameter tuning framework for black-box 3D LiDAR odometry algorithms. The proposed framework comprises offline parameter-error function modeling and online adaptive parameter selection. In the offline step, we run the odometry estimation algorithm for tuning with different parameters and environments and evaluate the accuracy of the estimated trajectories to build a surrogate function that predicts the trajectory estimation error for the given parameters and environments. Subsequently, we select the parameter set that is expected to result in good accuracy in the given environment based on trajectory error prediction with the surrogate function. The proposed framework does not require detailed information on the inner working of the algorithm to be tuned, and improves its accuracy by adaptively optimizing the parameter set. We first demonstrate the role of the proposed framework in improving the accuracy of odometry estimation across different environments with a simulation-based toy example. Further, an evaluation on the public dataset KITTI shows that the proposed framework can improve the accuracy of several odometry estimation algorithms in practical situations.

Adaptive control can address model uncertainty in control systems. However, it is preliminarily designed for tracking control. Recent advancements in the control of quadruped robots show that force control can effectively realize agile and robust locomotion. In this paper, we present a novel adaptive force-based control framework for legged robots. We introduce a new architecture in our proposed approach to incorporate adaptive control into quadratic programming (QP) force control. Since our approach is based on force control, it also retains the advantages of the baseline framework, such as robustness to uneven terrain, controllable friction constraints, or soft impacts. Our method is successfully validated in both simulation and hardware experiments. While the baseline QP control has shown a significant degradation in the body tracking error with a small load, our proposed adaptive force-based control can enable the 12-kg Unitree A1 robot to walk on rough terrains while carrying a heavy load of up to 6 kg (50% of the robot weight). When standing with four legs, our proposed adaptive control can even allow the robot to carry up to 11 kg of load (92% of the robot weight) with less than 5-cm tracking error in the robot height.

This paper proposes a novel active visuo-tactile based methodology wherein the accurate estimation of the time-invariant SE(3) pose of objects is considered for autonomous robotic manipulators. The robot equipped with tactile sensors on the gripper is guided by a vision estimate to actively explore and localize the objects in the unknown workspace. The robot is capable of reasoning over multiple potential actions, and execute the action to maximize information gain to update the current belief of the object. We formulate the pose estimation process as a linear translation invariant quaternion filter (TIQF) by decoupling the estimation of translation and rotation and formulating the update and measurement model in linear form. We perform pose estimation sequentially on acquired measurements using very sparse point cloud (≤ 15 points) as acquiring each measurement using tactile sensing is time consuming. Furthermore, our proposed method is computationally efficient to perform an exhaustive uncertainty-based active touch selection strategy in real-time without the need for trading information gain with execution time. We evaluated the performance of our approach extensively in simulation and by a robotic system.

Recent visual pose estimation and tracking solutions provide notable results on popular datasets such as T-LESS and YCB. However, in the real world, we can find ambiguous objects that do not allow exact classification and detection from a single view. In this work, we propose a framework that, given a single view of an object, provides the coordinates of a next viewpoint to discriminate the object against similar ones, if any, and eliminates ambiguities. We also describe a complete pipeline from a real object’s scans to the viewpoint selection and classification. We validate our approach with a Franka Emika Panda robot and common household objects featured with ambiguities. We released the source code to reproduce our experiments.