ICRA 2021 Papers

Human–robot interaction (HRI) has been widely researched in diverse applications. A robot following a person is one such scenario investigated in the HRI field. However, human movements and actions are complex and can change dramatically. We herein demonstrate a machine learning-based system that allows a person-following robot to track in real-time the predicted future motion of a walking human, from a first-person perspective. We assume that a depth sensor that can detect the human skeleton is loaded on a mobile robot to provide data on the user’s motion from a first-person perspective. The system calculates the coordinates of the center of gravity (COG) and 25 body joints of the user. These coordinates of COG and 25 body joints are relative to the robot based on the position of the person tracked, and these are used for the input dataset of a neural network (NN) that predicts human motion. A five-layered NN estimates the relative vectors in real-time between the current person’s COG and the future position of the 25 body joints. Using a proportional–integral–derivative (PID) controller, the person-following robot can track the predicted position of a walking human 0.5 s in advance to increase the robustness of following and to avoid delays.

Visual navigation for autonomous agents is a core task in the fields of computer vision and robotics. Learning-based methods, such as deep reinforcement learning, have the potential to outperform the classical solutions developed for this task; however, they come at a significantly increased computational load. Through this work, we design a novel approach that focuses on performing better or comparable to the existing learning-based solutions but under a clear time/computational budget. To this end, we propose a method to encode vital scene semantics such as traversable paths, unexplored areas, and observed scene objects–alongside raw visual streams such as RGB, depth, and semantic segmentation masks—into a semantically informed, top-down egocentric map representation. Further, to enable the effective use of this information, we introduce a novel 2-D map attention mechanism, based on the successful multi-layer Transformer networks. We conduct experiments on 3-D reconstructed indoor PointGoal visual navigation and demonstrate the effectiveness of our approach. We show that by using our novel attention schema and auxiliary rewards to better utilize scene semantics, we outperform multiple baselines trained with only raw inputs or implicit semantic information while operating with an 80% decrease in the agent’s experience.

The rapid development of autonomous driving and mobile mapping calls for off-the-shelf LiDAR SLAM solutions that are adaptive to LiDARs of different specifications on various complex scenarios. To this end, we propose MULLS, an efficient, low-drift, and versatile 3D LiDAR SLAM system. For the front-end, roughly classified feature points (ground, facade, pillar, beam, etc.) are extracted from each frame using dual-threshold ground filtering and principal components analysis. Then the registration between the current frame and the local submap is accomplished efficiently by the proposed multi-metric linear least square iterative closest point algorithm. Point-to-point (plane, line) error metrics within each point class are jointly optimized with a linear approximation to estimate the ego-motion. Static feature points of the registered frame are appended into the local map to keep it updated. For the back-end, hierarchical pose graph optimization is conducted among regularly stored history submaps to reduce the drift resulting from dead reckoning. Extensive experiments are carried out on three datasets with more than 100,000 frames collected by seven types of LiDAR on various outdoor and indoor scenarios. On the KITTI benchmark, MULLS ranks among the top LiDAR-only SLAM systems with real-time performance.

Symmetric environment is one of the most intractable and challenging scenarios for mobile robots to accomplish global localization tasks, due to the highly similar geometrical structures and insufficient distinctive features. Existing localization solutions in such scenarios either depend on pre-deployed infrastructures which are expensive, inflexible, and hard to maintain; or rely on single sensor-based methods whose initialization module is incapable to provide enough unique information. Thus, this paper proposes a novel Multi-Sensor based Two-Step Localization framework named MSTSL, which addresses the problem of mobile robot global localization in geometrically symmetric environments by utilizing the measured magnetic field, 2-D LiDAR, and wheel odometry information. The proposed system mainly consists of two steps: 1) Magnetic Field-based Initialization, and 2) LiDAR-based Localization. Based on the pre-built magnetic field database, multiple initial hypotheses poses can firstly be determined by the proposed two-stage initialization algorithm. Then, utilizing the obtained multiple initial hypotheses, the robot can be localized more accurately by LiDAR-based localization. Extensive experiments demonstrate the practical utility and accuracy of the proposed system over the alternative approaches in real-world scenarios.

For large-scale tasks, coverage path planning (CPP) can benefit greatly from multiple robots. In this paper, we present an efficient algorithm MSTC∗ for multi-robot coverage path planning (mCPP) based on spiral spanning tree coverage (Spiral-STC). Our algorithm incorporates strict physical constraints like terrain traversability and material load capacity. We compare our algorithm against the state-of-the-art in mCPP for regular grid maps and real field terrains in simulation environments. The experimental results show that our method significantly outperforms existing spiral-STC based mCPP methods. Our algorithm can find a set of well-balanced workload distributions for all robots and therefore, achieve the overall minimum time to complete the coverage.

Gaussian Process (GP) formulation of continuous-time trajectory offers a fast solution to the motion planning problem via probabilistic inference on factor graph. However, often the solution converges to in-feasible local minima and the planned trajectory is not collision-free. We propose a message passing algorithm that is more sensitive to obstacles with fast convergence time. We leverage the utility of min-sum message passing algorithm that performs local computations at each node to solve the inference problem on factor graph. We first introduce the notion of compound factor node to transform the factor graph to a linearly structured graph. We next develop an algorithm denoted as Min-sum Message Passing algorithm for Motion Planning (MS2MP) that combines numerical optimization with message passing to find collision- free trajectories. MS2MP performs numerical optimization to solve non-linear least square minimization problem at each compound factor node and then exploits the linear structure of factor graph to compute the maximum a posteriori (MAP) estimation of complete graph by passing messages among graph nodes. The decentralized optimization approach of each compound node increases sensitivity towards avoiding obstacles for harder planning problems. We evaluate our algorithm by performing extensive experiments for exemplary motion planning tasks for a robot manipulator. Our evaluation reveals that MS2MP improves existing work in convergence time and success rate.

Neural Architecture Search (NAS) has proved effective in offering outperforming alternatives to handcrafted neural networks. In this paper we analyse the benefits of NAS for image classification tasks under strict computational constraints. Our aim is to automate the design of highly efficient deep neural networks, capable of offering fast and accurate predictions and that could be deployed on a low-memory, low-power system-on-chip. The task thus becomes a three-party trade-off between accuracy, computational complexity, and memory requirements. To address this concern, we propose Multi-Scale Resource-Aware Neural Architecture Search (MS-RANAS). We employ a one-shot architecture search approach in order to obtain a reduced search cost and we focus on an anytime prediction setting. Through the usage of multiple-scaled features and early classifiers, we achieved state-of-the-art results in terms of accuracy-speed trade-off.

In multi-agent applications such as surveillance and logistics, fleets of mobile agents are often expected to coordinate and safely visit a large number of goal locations as efficiently as possible. The multi-agent planning problem in these applications involves allocating and sequencing goals for each agent while simultaneously producing conflict-free paths for the agents. In this article, we introduce a new algorithm called MS* which computes an optimal solution for this multi-agent problem by fusing and advancing state of the art solvers for multi-agent path finding (MAPF) and multiple travelling salesman problem (mTSP). MS* leverages our prior subdimensional expansion approach for MAPF and embeds the mTSP solvers to optimally allocate and sequence goals for agents. Numerical results show that our new algorithm can solve the multi-agent problem with 20 agents and 50 goals in a minute of CPU time on a standard laptop.

Local planning is one of the key technologies for mobile robots to achieve full autonomy and has been widely investigated. To evaluate mobile robot local planning approaches in a unified and comprehensive way, a mobile robot local planning benchmark called MRPB 1.0 is newly proposed in this paper. The benchmark facilitates both motion planning researchers who want to compare the performance of a new local planner relative to many other state-of-the-art approaches as well as end users in the mobile robotics industry who want to select a local planner that performs best on some problems of interest. We elaborately design various simulation scenarios to challenge the applicability of local planners, including large-scale, partially unknown, and dynamic complex environments. Furthermore, three types of principled evaluation metrics are carefully designed to quantitatively evaluate the performance of local planners, wherein the safety, efficiency, and smoothness of motions are comprehensively considered. We present the application of the proposed benchmark in two popular open-source local planners to show the practicality of the benchmark. In addition, some insights and guidelines about the design and selection of local planners are also provided. The benchmark website [1] contains all data of the designed simulation scenarios, detailed descriptions of these scenarios, and example code.

In this work, we propose a global planning strategy specifically designed for precision agriculture settings, where field activities may have different requirements ranging from a full orchard inspection to sparse targeted per-plant interventions. This global planning strategy is formulated as a novel Multi-Platform Steiner Traveling Salesman Problem (MP-STSP) where, in order to guarantee the exploitation of multiple moving platforms and the minimization of the overall operational time, the proposed formulation explicitly takes into account the time required to perform each task. By doing so, the computed itineraries attempt to balance the workload among the deployed platforms. Comparative simulations, inspired by the needs of the EU H2020 Project PANTHEON 1, are provided to numerically demonstrate the effectiveness of the proposed planning strategy for an orchard precision agriculture setting.

Robot design is a time-consuming process involving repeated experiments in a variety of environments to optimize multiple, possibly conflicting performance metrics. Moreover, the optimal robot performance for a given design depends on how the robot adapts its behavior to its environment. We propose a multi-objective Bilevel Bayesian optimization (MO-BBO) technique to automate the process of form-behavior co-design. The approach expands the Pareto front of multiple metrics by simultaneously exploring the robot design and behavior. MO-BBO uses a bilevel optimization of the acquisition function with design and behavior parameters being the high- and low-level decision variables, respectively. In the low-level, we always choose environment-aware behaviors that maximize each metric. We evaluate MO-BBO in applications to grasping gripper design and bimanual arm placement, and show that our method can efficiently focus samples on the Pareto front and generate a diversity of designs.

Autonomous navigation in crowded, complex urban environments requires interacting with other agents on the road. A common solution to this problem is to use a prediction model to guess the likely future actions of other agents. While this is reasonable, it leads to overly conservative plans because it does not explicitly model the mutual influence of the actions of interacting agents. This paper builds a reinforcement learning-based method named MIDAS where an Ego agent learns to affect the control actions of other cars in urban driving scenarios. MIDAS uses an attention mechanism to handle an arbitrary number of other agents and includes a "driver-type" parameter to learn a single policy that works across different planning objectives. We build a simulation environment that enables diverse interaction experiments with a large number of agents and develop methods for quantitatively studying the safety, efficiency, and interaction among vehicles. MIDAS is validated using extensive experiments and we show that it (i) can work across different road geometries, (ii) results in an adaptive Ego policy that can be tuned easily to satisfy different performance criteria, such as aggressive or cautious driving, (iii) is robust to changes in the driving policies of external agents, and (iv) is safer and more efficient than existing approaches to interaction-aware decision-making. Code available here.

6D pose estimation of objects is an important part of robot grasping. The latest research trend on 6D pose estimation is to train a deep neural network to directly predict the 2D projection position of the 3D key points from the image, establish the corresponding relationship, and finally use Pespective-n-Point (PnP) algorithm performs pose estimation. The current challenge of pose estimation is that when the object texture-less, occluded and scene clutter, the detection accuracy will be reduced, and most of the existing algorithm models are large and cannot take the real-time requirements. In this paper, we introduce a Multi-directional Feature Pyramid Network, MFPN, which can efficiently integrate and utilize features. We combined the Cross Stage Partial Network (CSPNet) with MFPN to design a new network for 6D pose estimation, MFPN-6D. At the same time, we propose a new confidence calculation method for object pose estimation, which can fully consider spatial information and plane information. At last, we tested our method on the LINEMOD and Occluded-LINEMOD datasets. The experimental results demonstrate that our algorithm is robust to textureless materials and occlusion, while running more efficiently compared to other methods.

Depth completion aims to recover the dense depth map from sparse depth data and RGB image respectively. However, due to the huge difference between the multi-modal signal input, vanilla convolutional neural network and simple fusion strategy cannot extract features from sparse data and aggregate multi-modal information effectively. To tackle this problem, we design a novel network architecture that takes full advantage of multi-modal features for depth completion. An effective Pre-completion algorithm is first put forward to increase the density of the input depth map and to provide distribution priors. Moreover, to effectively fuse the image features and the depth features, we propose a multi-modal deep aggregation block that consists of multiple connection and aggregation pathways for deeper fusion. Furthermore, based on the intuition that semantic image features are beneficial for accurate contour, we introduce the deformable guided fusion layer to guide the generation of the dense depth map. The resulting architecture, called MDANet, outperforms all the stateof-the-art methods on the popular KITTI Depth Completion Benchmark, meanwhile with fewer parameters than recent methods. The code of this work will be available at https://github.com/USTC-Keyanjie/MDANet_ICRA2021.

Occupancy grids have been widely used for mapping with mobile robots for several decades. Occupancy grids discretize the analog environment and seek to determine the occupancy probability of each cell. More recent occupancy grid mapping algorithms have shown the advantage of capturing cell correlations in the measurement model and the posterior. By estimating the probability of a given map as opposed to a cell, these algorithms have been able to better capture the occupancy probability of cells in the map. The advantage of incorporating data-driven prior probabilities in occupancy grid mapping is explored. A form of Markov Chain Monte Carlo (MCMC) known as Gibbs sampling allows us to sample maps from the full posterior. Previous research has sampled the occupancy probability of each cell, but this paper extends that work to sample a larger patch of cells and highlights the benefit of obtaining the prior for each patch from real maps.

Traditional multi-robot motion planning (MMP) focuses on computing trajectories for multiple robots acting in an environment, such that the robots do not collide when the trajectories are taken simultaneously. In safety-critical applications, a human supervisor may want to verify that the plan is indeed collision-free. In this work, we propose a notion of explanation for a plan of MMP, based on visualization of the plan as a short sequence of images representing time segments, where in each time segment the trajectories of the agents are disjoint, clearly illustrating the safety of the plan. We show that standard notions of optimality (e.g., makespan) may create conflict with short explanations. Thus, we propose meta-algorithms, namely multi-agent plan segmenting-X (MAPS-X) and its lazy variant, that can be plugged on existing centralized sampling-based tree planners X to produce plans with good explanations using a desirable number of images. We demonstrate the efficacy of this explanation-planning scheme and extensively evaluate the performance of MAPS-X and its lazy variant in various environments and agent dynamics.

This paper introduces a leg-wheel transformable quadruped robot named Lywal which can switch to the leg-mode and the wheel-mode for locomotion, and the claw-mode for picking up and transport functions. First, the mechanical structure of Lywal is designed by using an innovative 2-DoF transformable mechanism. Second, the calculation of kinematics is analyzed in detail. Then, the switching-mode strategy and the mobile control strategies in different modes are designed. Finally, the prototype of Lywal is built. The properties of the mobile modes are analyzed, and the picking-up and transport functions of the claw-mode are verified through physical experiments.

Parametric models that represent layout in terms of scene attributes are an attractive avenue for road scene understanding in autonomous navigation. Prior works that rely only on ground imagery are limited by the narrow field of view of the camera, occlusions and perspective foreshortening. In this paper, we demonstrate the effectiveness of using aerial imagery as an additional modality to overcome the above challenges. We propose a novel architecture, Unified, that combines features from both aerial and ground imagery to infer scene attributes. We quantitatively evaluate on the KITTI dataset and show that our Unified model outperforms prior works. Since this dataset is limited to road scenes close to the vehicle, we supplement the publicly available Argoverse dataset with scene attribute annotations and evaluate on far-away scenes. We show both quantitatively and qualitatively, the importance of aerial imagery in understanding road scenes, especially in regions farther away from the ego-vehicle. All code, models, and data, including scene attribute annotations on the Argoverse dataset along with collected and processed aerial imagery, are available at https://bit.ly/2QsKNeR.

To function autonomously in the physical world, humanoid robots need high-fidelity sensing systems, especially for forces that cannot be easily modeled. Modeling forces in robot feet is particularly challenging due to static indeterminacy, thereby requiring direct sensing. Unfortunately, resolving forces in the feet of some smaller-sized humanoids is limited both by the quality of sensors and the current algorithms used to interpret the data. This paper presents light-weight, low-cost and open-source force-sensing shoes to improve force measurement for popular smaller-sized humanoid robots, and a method for calibrating the shoes. The shoes measure center of pressure (CoP) and normal ground reaction force (GRF). The calibration method enables each individual shoe to reach high measurement precision by applying known forces at different locations of the shoe and using a regularized least squares optimization to interpret sensor outputs. A NAOTM robot is used as our experimental platform. Experiments are conducted to compare the measurement performance between the shoes and the robot’s factory-installed force-sensing resistors (FSRs), and to evaluate the calibration method over these two sensing modules. Experimental results show that the shoes significantly improve CoP and GRF measurement precision compared to the robot’s built-in FSRs. Moreover, the developed calibration method improves the measurement performance for both our shoes and the built-in FSRs.