ICRA 2022 论文列表

In this paper, a Deep Neural Network is trained using Reinforcement Learning in order to drift on arbitrary trajectories which are defined by a sequence of waypoints. In a first step, a highly accurate vehicle simulation is used for the training process. Then, the obtained policy is refined and validated on a self-built model car. The chosen reward function is inspired by the scoring process of real life drifting competitions. It is kept simple and thus applicable to very general scenarios. The experimental results demonstrate that a relatively small network, given only a few measurements and control inputs, already achieves an outstanding performance. In simulation, the learned controller is able to reliably hold a steady state drift. Moreover, it is capable of generalizing to arbitrary, previously unknown trajectories and different driving conditions. After transferring the learned controller to the model car, it also performs surprisingly well given the physical constraints.

Maneuverable multicamera systems offer potential benefits in abdominal minimally-invasive procedures, including multi-view scene reconstruction and optimal viewpoint capture. Effective autonomous movement and re-positioning of such systems, however, remains an open challenge due to dynamic motion constraints, deforming surgical scenes, and visual artifacts such as motion blur, specular reflections, and blood stains [1]. Despite these existing roadblocks, multicamera systems have been used both to provide surgeons with stable and analytically optimized viewpoints [2] and to enable 3D surgical scene reconstruction [3] that directly contributes towards the possibility of task autonomy [4] and AR-enhanced surgical procedures [5]. These methods, however, often require extensive, high-dimensional continuous data sets, and may not value scene discovery. To that end, this project presents a novel curiosity driven multicamera viewpoint adjustment framework, Ac, that aims to simultaneously (a) explore and maximize weighted 3D reconstructable coverage; (b) limit unnecessary camera motion; and (c) relieve data-intensiveness through dimension-reduced state representations. The developed algorithms are comparatively evaluated against three baseline methods on simulated surgical sequences, and results demonstrate performance enhancements with the presented methods.

Covariance estimation for the Iterative Closest Point (ICP) point cloud registration algorithm is essential for state estimation and sensor fusion purposes. We argue that a major source of error for ICP is in the input data itself, from the sensor noise to the scene geometry. Benefiting from recent developments in deep learning for point clouds, we propose a data-driven approach to learn an error model for ICP. We estimate covariances modeling data-dependent heteroscedastic aleatoric uncertainty, and epistemic uncertainty using a variational Bayesian approach. The system evaluation is performed on LiDAR odometry on different datasets, highlighting good results in comparison to the state of the art.

As a basic part of the centroidal dynamics, an-gular momentum plays a critical role in humanoid motion control. Therefore, how to explicitly express and control an-gular momentum through whole-body motion is an important topic for researchers. This study discusses the selection of the generalized velocity corresponding to whole-body angular momentum. Based on the discussion, we present a method that decouples the inertia effect in centroidal angular momentum and applies it in resolved viscoelasticity control, which achieves the angular momentum control by whole-body compliance in an interpretable way without complicated calculation. At last, we validate the feasibility and effectiveness of proposed method in forward dynamics simulation of balancing control in the double and single support states and landing motion after hopping.

Ride-sharing has important implications for improving the efficiency of mobility-on-demand systems. However, it remains a challenge due to the complex dynamics between vehicles and requests. This paper presents a decentralized ride-sharing algorithm suitable for shared autonomous vehicles (SAVs) deployment. The ride-sharing problem is formulated as a multi-agent reinforcement learning problem. We explore state representation with the request-vehicle graph to encode shareability and potential coordination information. We use a graph attention network to build a hierarchical structure that unifies ride-sharing assignments with rebalancing and handles real-world scenarios where hundreds of user requests can be associated with vehicles. We show results in both generic grid-world and SUMO simulation with real-world data from the Manhattan area. We empirically demonstrate that our proposed approach can achieve similar performance compared with a state-of-the-art centralized optimization method and higher computation efficiency.

In this paper we analyze the equilibrium points of a collaborative transportation task, composed of two unmanned aerial vehicles and a payload - in this case, a bar. Moreover, centralized and decentralized linear model predictive controllers are designed, where the nonlinear dynamics are linearized around the equilibrium points previously analyzed. A comparison between the centralized and decentralized formulations is provided, based on experimental results for both setups, and considering the time to solution and performance of each controller. Our findings provide new operational equilibrium points that can be paired with predictive model-based controllers for efficient operation.

The problem of multi-robot navigation of connectivity maintenance is challenging in multi-robot applications. This work investigates how to navigate a multi-robot team in unknown environments while maintaining connectivity. We propose a reinforcement learning (RL) approach to develop a decentralized policy, which is shared among multiple robots. Given range sensor measurements and the positions of other robots, the policy aims to generate control commands for navigation and preserve the global connectivity of the robot team. We incorporate connectivity concerns into the RL framework as constraints and introduce behavior cloning to reduce the exploration complexity of policy optimization. The policy is optimized with all transition data collected by multiple robots in random simulated scenarios. We validate the effectiveness of the proposed approach by comparing different combinations of connectivity constraints and behavior cloning. We also show that our policy can generalize to unseen scenarios in both simulation and holonomic robots experiments.

Point clouds from 3D light detection and ranging (LiDAR) are widely used. Noise caused by falling snow reduces the availability of point clouds. Due to the sparseness of LiDAR point clouds and the fact that the snow point clouds are easily affected by multi factors such as wind or snowfall conditions, it is difficult to accurately remove the snow while preserving the details of the point clouds. To solve the problem, this paper presents a de-snowing approach combining the intensity and spatial-temporal features. An intensity-based filter firstly removes the snow. Then a repairing method restores the non-snow points based on the spatial-temporal features. Experimental results demonstrate that our approach outperforms existing work in the literature and performs the least damage to the point clouds in different snowfall scenarios.

Grasping made impressive progress during the last few years thanks to deep learning. However, there are many objects for which it is not possible to choose a grasp by only looking at an RGB-D image, might it be for physical reasons (e.g., a hammer with uneven mass distribution) or task constraints (e.g., food that should not be spoiled). In such situations, the preferences of experts need to be taken into account. In this paper, we introduce a data-efficient grasping pipeline (Latent Space GP Selector - LGPS) that learns grasp prefer-ences with only a few labels per object (typically 1 to 4) and generalizes to new views of this object. Our pipeline is based on learning a latent space of grasps with a dataset generated with any state-of-the-art grasp generator (e.g., Dex-Net). This latent space is then used as a low-dimensional input for a Gaussian process classifier that selects the preferred grasp among those proposed by the generator. The results show that our method outperforms both GR-ConvNet and GG-CNN (two state-of-the-art methods that are also based on labeled grasps) on the Cornell dataset, especially when only a few labels are used: only 80 labels are enough to correctly choose 80% of the grasps (885 scenes, 244 objects). Results are similar on our dataset (91 scenes, 28 objects).

This paper presents four data-driven system models for a magnetically controlled swimmer. The models were derived directly from experimental data, and the accuracy of the models was experimentally demonstrated. Our previous study successfully implemented two non-model-based control algorithms for 3D path-following using PID and model reference adaptive controller (MRAC). This paper focuses on system identification using only experimental data and a model-based control strategy. Four system models were derived: (1) a physical estimation model, (2, 3) Sparse Identification of Nonlinear Dynamics (SINDY), linear system and nonlinear system, and (4) multilayer perceptron (MLP). All four system models were implemented as an estimator of a multi-step Kalman filter. The maximum required sensing interval was increased from 180 ms to 420 ms and the respective tracking error decreased from 9 mm to 4.6 mm. Finally, a Model Predictive Controller (MPC) implementing the linear SINDY model was tested for 3D path-following and shown to be computationally efficient and offers performances comparable to other control methods.

Music to dance for humanoid robots is an interesting task. Robot dance generation is challenging when considering music pieces, human dancer motions, and robot stability simultaneously. Previous methods rely on human-designed motion library or stability constraints for robot postures. Hence, dance generation for humanoid robots requires expert design, which can be time-consuming across different humanoid platforms. In this work, we propose a novel method called DanceHAT, which generates stable humanoid dances by imitating human dancers with self-learning. DanceHAT is an adversarial training framework, which incorporates similarity loss and stability loss simultaneously. Furthermore, DanceHAT does not require human-designed features or robot model information. Experiments in the simulation environment and on the real robot demonstrate that our model can generate stable, diverse, and human-like dances for humanoid robots automatically. In addition, DanceHAT is a general training approach for robot imitation tasks with stability constraints, thus can be utilized in other humanoid tasks and will be researched in future works.

Visual servoing enables robotic systems to perform accurate closed-loop control, which is required in many applications. However, existing methods require either precise calibration of the robot kinematic model and cameras or use neural architectures that require large amounts of data to train. In this work, we present a method for unsupervised learning of visual servoing that does not require any prior calibration and is extremely data-efficient. Our key insight is that visual servoing does not depend on identifying the veridical kinematic and camera parameters, but instead only on an accurate generative model of image feature observations from the joint positions of the robot. We demonstrate that with our model architecture and learning algorithm, we can consistently learn accurate models from less than 50 training samples (which amounts to less than 1 min of unsupervised data collection), and that such data-efficient learning is not possible with standard neural architectures. Further, we show that by using the generative model in the loop and learning online, we can enable a robotic system to recover from calibration errors and to detect and quickly adapt to possibly unexpected changes in the robot-camera system (e.g. bumped camera, new objects).

Environment modeling is the backbone of how autonomous agents understand the world, and therefore has significant implications for decision-making and verification. Motivated by the success of relational mapping tools such as Lanelet2, we present the Dynamic Relation Graph (DRG). The DRG is a novel method for extending prior relational maps to include online observations, creating a unified en-vironment model which incorporates both prior and online data sources. Our prototype implementation models a finite set of heterogeneous features including road signage and pedestrian movement. However, the methodology behind the DRG can be expanded to a wider range of features in a fashion that does not increase the complexity of behavioral planning. Simulated stress tests indicate the DRG's effectiveness in decreasing decision-making complexity, and deployment on the University of Waterloo's WATonomous research vehicle demonstrates its practical utility. The prototype code will be released at github.com/WATonomous/DRG.

Underwater thermocline, common in the lakes and ocean, plays a vital role in meteorological forecasting in the ocean and lakes dynamics research. This letter proposes a method for rapid and multipoint observation of thermocline variations with time and space using an airdropped micro-profiler array, named the DRAGONFLY system. It comprises specially designed disposable low-cost micro-profilers, a general unmanned aerial carrier platform, and a ground control system. This system can conduct periodic profile observations at a single point or quickly survey a large area. A series of experiments to characterize the micro-profiler and the DRAGONFLY system were conducted in Qiandao Lake, China. We demonstrate the developed system with data from field experiments, which show very high flexibility, and feasibility to observe the lake thermocline, implying potential applications in ocean transient phenomena observation.

Safe UAV navigation is challenging due to the complex environment structures, dynamic obstacles, and uncertainties from measurement noises and unpredictable moving obstacle behaviors. Although plenty of recent works achieve safe navigation in complex static environments with sophisticated mapping algorithms, such as occupancy map and ESDF map, these methods cannot reliably handle dynamic environments due to the mapping limitation from moving obstacles. To address the limitation, this paper proposes a trajectory planning framework to achieve safe navigation considering complex static environments with dynamic obstacles. To reliably handle dynamic obstacles, we divide the environment representation into static mapping and dynamic object representation, which can be obtained from computer vision methods. Our framework first generates a static trajectory based on the proposed iterative corridor shrinking algorithm. Then, reactive chance-constrained model predictive control with temporal goal tracking is applied to avoid dynamic obstacles with uncertainties. The simulation results in various environments demonstrate the ability of our algorithm to navigate safely in complex static environments with dynamic obstacles.

Exploration of unknown environments is an important challenge in the field of robotics. While a single robot can achieve this task alone, evidence suggests it could be accomplished more efficiently by groups of robots, with advantages in terms of terrain coverage as well as robustness to failures. Exploration can be guided through belief maps, which provide probabilistic information about which part of the terrain is interesting to explore (either based on risk management or reward). This process can be centrally coordinated by building a collective belief map on a common server. However, relying on a central processing station creates a communication bottleneck and single point of failure for the system. In this paper, we present Distributed Online Risk-Aware (DORA) Explorer, an exploration system that leverages decentralized information sharing to update a common risk belief map. DORA-Explorer allows a group of robots to explore an unknown environment discretized as a 2D grid with obstacles, with high coverage while minimizing exposure to risk, effectively reducing robot failures.

Keypoint extraction including both keypoint detection and description is a fundamental step in a wide range of geometric multimedia applications. In recent years, many learning-based approaches for keypoint extraction emerge and achieve promising results. However, they usually design network architectures empirically and lack of considerations about the comprehensive performance, which leads to limited applications. In this paper, we propose a practical framework based on Neural Architecture Search (NAS) technology, DKNAS, which can search architectures automatically and maintain efficiency and effectiveness, simultaneously. To the best of our knowledge, the proposed framework is the first NAS framework for keypoint extraction. The evaluation on HPatches dataset shows that our method achieves state-of-the-art results in the metrics of repeatability, localization error, homography accuracy and matching scores. Besides, our model is applied to a traditional Simultaneous Localization and Mapping (SLAM) system, ORB-SLAM2, to replace the handcrafted keypoints. Experimental results demonstrate that the system adopting our model outperforms ORB-SLAM2 and some other deep keypoints enhanced systems.

We present a novel real-time visual odometry framework for a stereo setup of a depth and high-resolution event camera. Our framework balances accuracy and robustness against computational efficiency towards strong performance in challenging scenarios. We extend conventional edge-based semi-dense visual odometry towards time-surface maps obtained from event streams. Semi-dense depth maps are generated by warping the corresponding depth values of the extrinsically calibrated depth camera. The tracking module updates the camera pose through efficient, geometric semi-dense 3D-2D edge alignment. Our approach is validated on both public and self-collected datasets captured under various conditions. We show that the proposed method performs comparable to state-of-the-art RGB-D camera-based alternatives in regular conditions, and eventually outperforms in challenging conditions such as high dynamics or low illumination.

Automotive mmWave radar has been widely used in the automotive industry due to its small size, low cost, and complementary advantages to optical sensors (e.g., cameras, LiDAR, etc.) in adverse weathers, e.g., fog, raining, and snowing. On the other side, its large wavelength also poses fundamental challenges to perceive the environment. Recent advances have made breakthroughs on its inherent drawbacks, i.e., the multipath reflection and the sparsity of mmWave radar's point clouds. However, the frequency-modulated continuous wave modulation of radar signals makes it more sensitive to vehicles’ mobility than optical sensors. This work focuses on the problem of frequency shift, i.e., the Doppler effect distorts the radar ranging measurements and its knock-on effect on metric localization. We propose a new radar-based metric localization framework, termed DC-Loc, which can obtain more accurate location estimation by restoring the Doppler distortion. Specifically, we first design a new algorithm that explicitly compensates the Doppler distortion of radar scans and then model the measurement uncertainty of the Doppler-compensated point cloud to further optimize the metric localization. Extensive experiments using the public nuScenes dataset and CARLA simulator demonstrate that our method outperforms the state-of-the-art approach by 25.2% and 5.6% improvements in terms of translation and rotation errors, respectively.