ICRA 2021 论文列表

This paper presents a navigation network based deep reinforcement learning framework for autonomous indoor robot exploration. The presented method features a pattern cognitive non-myopic exploration strategy that can better reflect universal preferences for structure. We propose the Extendable Navigation Network (ENN) to encode the partially observed high-dimensional indoor Euclidean space to a sparse graph representation. The robot’s motion is generated by a learned Q-network whose input is the ENN. The proposed framework is applied to a robot equipped with a 2D LIDAR sensor in the GAZEBO simulation where floor plans of real buildings are implemented. The experiments demonstrate the efficiency of the framework in terms of exploration time.

This paper describes a novel framework for autonomous exploration in large and complex environments. We show that the framework is efficient as a result of its hierarchical structure, where at one level it maintains a sparse representation of the environment and at another level, a dense representation is used within a local planning horizon around the robot. The exploration path is computed at the two levels, coarsely at the global scale and finely around the robot. Such a framework produces detailed paths in the vicinity of the robot, while trades off data resolution far away from the robot for computational efficiency. In experiments, we evaluate our method with a real robot exploring large and complex indoor and outdoor environments. Results show that our method is twice as efficient in covering spaces while using less than one-fifth of processing in comparison to state-of-the-art methods.

To accurately predict future positions of different agents in traffic scenarios is crucial for safely deploying intelligent autonomous systems in the real-world environment. However, it remains a challenge due to the behavior of a target agent being affected by other agents dynamically and there being more than one socially possible paths the agent could take. In this paper, we propose a novel framework, named Dynamic Context Encoder Network (DCENet). In our framework, first, the spatial context between agents is explored by using self-attention architectures. Then, the two-stream encoders are trained to learn temporal context between steps by taking the respective observed trajectories and the extracted dynamic spatial context as input. The spatial-temporal context is encoded into a latent space using a Conditional Variational Auto-Encoder (CVAE) module. Finally, a set of future trajectories for each agent is predicted conditioned on the learned spatial-temporal context by sampling from the latent space, repeatedly. DCENet is evaluated on one of the most popular challenging benchmarks for trajectory forecasting Trajnet and reports a new state-of-the-art performance. It also demonstrates superior performance evaluated on the benchmark inD for mixed traffic at intersections. A series of ablation studies is conducted to validate the effectiveness of each proposed module. Our code is available at https://github.com/wtliao/DCENet.

We present a Multi-Criteria Decision Making (MCDM) framework specifically designed for planetary exploration. Our work is based on PROMETHEE II, which allows operators to add task-specific criteria and conditions. We extended this algorithm to improve its resource usage by reducing the number of candidate exploration goals that have to be evaluated and compared. This is crucial when given a large number of goals, as is typical for outdoor environments. In addition, we identified five different criteria for planetary exploration, including a novel criterion that we call Direction of Interest (DOI), and use a categorization of these for further resource optimizations. We thereby ensure that the CPU usage of our decision making method can meet the limited budget of a space rover. We present simulated and real-world experiments with the Lightweight Rover Unit (LRU) and show a reduction of the processing time for decision making of approx. 70%.

The big potential of humanoid robots is not restricted to the ground, but these versatile machines can be successfully employed in unconventional scenarios, e.g. space, where contacts are not always present. In these situations, the robot’s limbs can be used to assist or even generate the angular motion of the floating base, as a consequence of the centroidal momentum conservation. In this paper, we propose to combine, in the same whole-body motion control, visual servoing and centroidal momentum conservation. The former dictates a rotation to the floating humanoid to achieve a task in the Cartesian space; the latter is exploited to realize the desired rotation by moving the robot’s articulations. Simulations in a space scenario are carried out using COMAN, a humanoid robot developed at the Istituto Italiano di Tecnologia.

As urban environments manifest high levels of complexity it is of vital importance that safety systems embedded within autonomous vehicles (AVs) are able to accurately anticipate short-term future motion of nearby agents. This problem can be further understood as generating a sequence of coordinates describing the future motion of the tracked agent. Various proposed approaches demonstrate significant benefits of using a rasterised top-down image of the road, with a combination of Convolutional Neural Networks (CNNs), for extraction of relevant features that define the road structure (eg. driveable areas, lanes, walkways). In contrast, this paper explores use of Capsule Networks (CapsNets) in the context of learning a hierarchical representation of sparse semantic layers corresponding to small regions of the High-Definition (HD) map. Each region of the map is dismantled into separate geometrical layers that are extracted with respect to the agent’s current position. By using an architecture based on CapsNets the model is able to retain hierarchical relationships between detected features within images whilst also preventing loss of spatial data often caused by the pooling operation. We train and evaluate our model on publicly available dataset nuTonomy scenes and compare it to recently published methods. We show that our model achieves significant improvement over recently published works on deterministic prediction, whilst drastically reducing the overall size of the network.

Multicopters with collision-resilient designs can operate with trajectories involving collisions. This paper presents a sampling-based method that can exploit collisions for better motion planning. The method is built upon the basis of the RRT* algorithm and takes advantages of fast motion primitive generation and collision checking for multicopters. It generates collision states by detecting potential intersections between motion primitives and obstacles, and connects these states with other sampled states to form collision-inclusive trajectories. We show that allowing collision helps improve the performance of the sampling-based planner in narrow spaces like tunnels. Finally, an experiment of tracking the trajectory generated by the collision-inclusive planner is presented.

Visual tracking is a fundamental capability for robots tasked with humans and environment interaction. However, state-of-the-art visual tracking methods are still prone to failures and are imprecise when applied to challenging stereos, and their results are generally confidence agonistic. These methods depend on an embedded deep learning model to provide deterministic features or regression maps. A deterministic output with low confidence can result in disastrous consequences and lacks evidence needed for subsequent operations. Moreover, training data ambiguities or noise in the observations (so-called data uncertainty) can also lead to inherent uncertainty. In this paper, we focus on exploiting probabilistic Siamese visual tracking with a conditional variational autoencoder (CVAE). First, we build a bridge between the Siamese architecture and the CVAE and propose a novel Bayesian visual tracking method. Second, the proposed method generates a complete probability distribution that enables the production of multiple plausible tracking outputs. Third, CVAE conditioned by ground truth data encodes a low-dimensional latent space and conducts noise-injection training to prevent overfitting. Our proposed tracking method outperformed the state-of-the-art trackers on the VOT2016, VOT2018 and TColor-128 datasets.

We propose simple yet effective improvements in point representations and local neighborhood graph construction within the general framework of graph neural networks (GNNs) for 3D point cloud processing. As a first contribution, we propose to augment the vertex representations with important local geometric information of the points, followed by nonlinear projection using a MLP. As a second contribution, we propose to improve the graph construction for GNNs for 3D point clouds. The existing methods work with a k-NN based approach for constructing the local neighborhood graph. We argue that it might lead to reduction in coverage in case of dense sampling by sensors in some regions of the scene. The proposed methods aims to counter such problems and improve coverage in such cases. As the traditional GNNs were designed to work with general graphs, where vertices may have no geometric interpretations, we see both our proposals as augmenting the general graphs to incorporate the geometric nature of 3D point clouds. While being simple, we demonstrate with multiple challenging benchmarks, with relatively clean CAD models, as well as with real world noisy scans, that the proposed method achieves state of the art results on benchmarks for 3D classification (ModelNet40) , part segmentation (ShapeNet) and semantic segmentation (Stanford 3D Indoor Scenes Dataset). We also show that the proposed network achieves faster training convergence, i.e. ∼ 40% less epochs for classification. The project details are available at https://siddharthsrivastava.github.io/publication/geomgcnn/

Shallow water environments pose daunting scenarios for the operation of Unmanned Underwater Vehicles (UUVs), due to significantly larger wave disturbances being present in comparison to a typical deep sea situation. Performing inspection and maintenance tasks at close quarters in these conditions requires reliable control methods robust to external disturbances, allowing accurate position and attitude control, an aspect which classical control methods are often lacking. Improved performance can be achieved through predictive control methods, however, these require accurate and time-efficient estimations of the hydrodynamic forces produced by the immediate ocean environment around the vehicle. Considering this, we present a low-order model for faster-than-real time estimation of the wave-induced hydrodynamic forces acting on a submerged vehicle in various sea state conditions. The model is thoroughly corroborated by experimental tests, performed using a Remotely Operated Vehicle (ROV) situated at shallow depth whilst subjected to realistic sea wave disturbances. Validation between simulations and the collected experimental data showed a maximum normalised mean error deviation of 0.16 and 0.27 for surge and heave forces respectively, and 0.34 for the pitching moment. This empirical evidence demonstrates that accurate predictions of wave-generated forces can be produced through low-order models at a speed suitable for incorporation within predictive control architectures.

One major challenge in Assembly Sequence Planning (ASP) for complex real-world CAD-scenarios is to find an appropriate disassembly path for each assembled part. Complex real-world scenes are characterized by a large installation space. There each part has many different possible disassembly paths that differ in length and clearance. However, due to tight packing in the installation space, these paths can contain narrow passages. Therefore a motion planner is needed that is able to globally search for a reasonable path and to locally overcome narrow passages. Moreover, since motion planning requests are executed in the ASP context over and over again for many parts, both for those that can be disassembled in the next step and for those that cannot be yet, the motion planner has to be reliably fast.We present a new rigid body motion planner, called Expansive Voronoi Tree (EVT), which is optimized for complex ASP scenarios. The EVT estimates a globally reasonable path using a General Voronoi Diagram of the complete scene. With a novel EST-based sampling strategy, which is the contribution of this paper, it then locally explores the environment along the estimated path. The EVT automatically adapts to different clearance situations. It passes wide environments quickly and samples densely at narrow passages.We compare our EVT to state of the art motion planners which use different sampling strategies on a real-world data set consisting of a large subset of a car. The experiments show that the EVT is reliably many times faster and delivers shorter paths.

A motion-planning method that can adapt to changes in the surrounding environment is proposed and evaluated. Automation of work is progressing in factories and distribution warehouses due to labor shortages. However, utilizing robots for transport operations in a distribution warehouse faces a problem; that is, tasks for setting up a robot, such as adjustment of acceleration for stabilization of the transportation operation, are time consuming. To solve that problem, we developed an "evolvable robot motion-planning method." The aim of this method is to reduce the preparation cost by allowing the robot to automatically learn the optimized acceleration according to the weight and center of gravity of the objects to be transported. It was experimentally demonstrated that the proposed method can learn the optimized acceleration control from time-series data such as sensor information. The proposed method was evaluated in a simulator environment, and the results of the evaluation demonstrate that the learned model reduced the inertial force due to the acceleration of robot motion and shortened the transport time by 35% compared with the conventional method of manual adjustment. The proposed method was also evaluated in a real machine environment, and the evaluation results demonstrate that the method can be applied to a real robot. Since the speed of the robot does not need to be adjusted in the case of the proposed method, the adjustment man-hours can be reduced.

Event-based vision sensors achieve up to three orders of magnitude better speed vs. power consumption trade off in high-speed control of UAVs compared to conventional image sensors. Event-based cameras produce a sparse stream of events that can be processed more efficiently and with a lower latency than images, enabling ultra-fast vision-driven control. Here, we explore how an event-based vision algorithm can be implemented as a spiking neuronal network on a neuromorphic chip and used in a drone controller. We show how seamless integration of event-based perception on chip leads to even faster control rates and lower latency. In addition, we demonstrate how online adaptation of the SNN controller can be realised using on-chip learning. Our spiking neuronal network on chip is the first example of a neuromorphic vision-based controller on chip solving a high-speed UAV control task. The excellent scalability of processing in neuromorphic hardware opens the possibility to solve more challenging visual tasks in the future and integrate visual perception in fast control loops.

Several applications of video rely on camera calibration, a key enabler towards the measurement of metric parameters from images. Existing calibration methods have not been adapted for use with cameras that are hard to reach such as security cameras. This paper presents a drone-based camera calibration technique for the calibration of cameras in a broader range of operating environments. The technique is enabled by the use of a drone that is equipped with location sensors. In the proposed approach, the drone is used to sample points in the 3D space and its detection on the images provides the 2D matching points enabling the calibration. To design a flight path that allows for an accurate calibration, the paper proposes a methodology to evaluate the impact of path parameters and drone localization and detection uncertainties on the calibration uncertainty. This methodology is applied to evaluate the performance of different sampling paths for the calibration of a large diversity of cameras for the purpose of recommending reliable path parameters.

Discriminative data representations for point cloud data are critical for computer vision applications. Recently, the Hierarchical Gaussian Mixture Model (HGMM) has become a popular representation due to its compactness and real-time execution. However, HGMM still lacks a well-designed and robust initialization criterion. Ad-hoc initializations for HGMM can lead to a low discriminative clustering capability, slow convergence, and loss of scale-invariance. To adopt the optimal initialization scheme, we evaluate four potential candidates: K-Means++, Fuzzy C-Means (FCM), uniform, and random initialization across a few synthetic and measured datasets. Our experiments involve comparing the quality of HGMM point cloud reconstruction based on different initialization methods. The reconstruction quality is evaluated by the peak signal-to-noise ratio (PSNR). Our experiments show that clustering-based initialization methods can result in higher-quality HGMMs because of i) faster convergence of the Expectation-Maximization (EM) optimization, ii) better scaleinvariance across differently sized datasets, and iii) greater stability for different initial scales of covariance matrices of the HGMM.

We present a method to estimate two-dimensional, time-invariant oceanic flow fields based on data from both ensemble forecasts and online measurements. Our method produces a realistic estimate in a computationally efficient manner suitable for use in marine robotics for path planning and related applications. We use kernel methods and singular value decomposition to find a compact model of the ensemble data that is represented as a linear combination of basis flow fields and that preserves the spatial correlations present in the data. Online measurements of ocean current, taken for example by marine robots, can then be incorporated using recursive Bayesian estimation. We provide computational analysis, performance comparisons with related methods, and demonstration with real-world ensemble data to show the computational efficiency and validity of our method. Possible applications in addition to path planning include active perception for model improvement through deliberate choice of measurement locations.

It is ubiquitously accepted that during the autonomous navigation of the quadrotors, one of the most widely adopted unmanned aerial vehicles (UAVs), safety always has the highest priority. However, it is observed that the ego airflow disturbance can be a significant adverse factor during flights, causing potential safety issues, especially in narrow and confined indoor environments. Therefore, we propose a novel method to estimate and adapt indoor ego airflow disturbance of quadrotors, meanwhile applying it to trajectory planning. Firstly, the hover experiments for different quadrotors are conducted against the proximity effects. Then with the collected acceleration variance, the disturbances are modeled for the quadrotors according to the proposed formulation. The disturbance model is also verified under hover conditions in different reconstructed complex environments. Furthermore, the approximation of Hamilton-Jacobi reachability analysis is performed according to the estimated disturbances to facilitate the safe trajectory planning, which consists of kinodynamic path search as well as B-spline trajectory optimization. The whole planning framework is validated on multiple quadrotor platforms in different indoor environments.

With growing access to versatile robotics, it is beneficial for end users to be able to teach robots tasks without needing to code a control policy. One possibility is to teach the robot through successful task executions. However, near-optimal demonstrations of a task can be difficult to provide and even successful demonstrations can fail to capture task aspects key to robust skill replication. Here, we propose a learning from demonstration (LfD) approach that enables learning of robust task definitions without the need for near-optimal demonstrations. We present a novel algorithmic framework for learning tasks based on the ergodic metric—a measure of information content in motion. Moreover, we make use of negative demonstrations—demonstrations of what not to do—and show that they can help compensate for imperfect demonstrations, reduce the number of demonstrations needed, and highlight crucial task elements improving robot performance. In a proof-of-concept example of cart-pole inversion, we show that negative demonstrations alone can be sufficient to successfully learn and recreate a skill. Through a human subject study with 24 participants, we show that consistently more information about a task can be captured from combined positive and negative (posneg) demonstrations than from the same amount of just positive demonstrations. Finally, we demonstrate our learning approach on simulated tasks of target reaching and table cleaning with a 7-DoF Franka arm. Our results point towards a future with robust, data-efficient LfD for novice users.

This paper presents a novel factor graph-based approach to solve the discrete-time finite-horizon Linear Quadratic Regulator problem subject to auxiliary linear equality constraints within and across time steps. We represent such optimal control problems using constrained factor graphs and optimize the factor graphs to obtain the optimal trajectory and the feedback control policies using the variable elimination algorithm with a modified Gram-Schmidt process. We prove that our approach has the same order of computational complexity as the state-of-the-art dynamic programming approach. Furthermore, current dynamic programming approaches can only handle equality constraints between variables at the same time step, but ours can handle equality constraints among any combination of variables at any time step while maintaining linear complexity with respect to trajectory length. Our approach can be used to efficiently generate trajectories and feedback control policies to achieve periodic motion or repetitive manipulation.