ICRA 2021 论文列表

We present an efficient, elastic 3D LiDAR reconstruction framework which can reconstruct up to maximum Li-DAR ranges (60 m) at multiple frames per second, thus enabling robot exploration in large-scale environments. Our approach only requires a CPU. We focus on three main challenges of large-scale reconstruction: integration of long-range LiDAR scans at high frequency, the capacity to deform the reconstruction after loop closures are detected, and scalability for long-duration exploration. Our system extends upon a state-of-the-art efficient RGB-D volumetric reconstruction technique, called supereight, to support LiDAR scans and a newly developed submapping technique to allow for dynamic correction of the 3D reconstruction. We then introduce a novel pose graph clustering and submap fusion feature to make the proposed system more scalable for large environments. We evaluate the performance using two public datasets including outdoor exploration with a handheld device and a drone, and with a mobile robot exploring an underground room network. Experimental results demonstrate that our system can reconstruct at 3 Hz with 60 m sensor range and ~5 cm resolution, while state-of-the-art approaches can only reconstruct to 25 cm resolution or 20 m range at the same frequency.

This paper explores the use of passive, stereo sensing for vision-based navigation. The traditional approach uses dense depth algorithms, which can be computationally costly or potentially inaccurate. These drawbacks compound when including the additional computational demands associated to the sensor fusion, collision checking, and path planning modules that interpret the dense depth measurements. These problems can be avoided through the use of the stixel representation, a compact and sparse visual representation for local free-space. When integrated into a Planning in Perception Space based hierarchical navigation framework, stixels permit fast and scalable navigation for different robot geometries. Computational studies quantify the processing performance and demonstrate the favorable scaling properties over comparable dense depth methods. Navigation benchmarking demonstrates more consistent performance across high and low performance compute hardware for PiPS-based stixel navigation versus traditional hierarchical navigation.

We propose an efficient way of solving optimal control problems for rigid-body systems on the basis of inverse dynamics and the multiple-shooting method. We treat all variables, including the state, acceleration, and control input torques, as optimization variables and treat the inverse dynamics as an equality constraint. We eliminate the update of the control input torques from the linear equation of Newton’s method by applying condensing for inverse dynamics. The size of the resultant linear equation is the same as that of the multiple-shooting method based on forward dynamics except for the variables related to the passive joints and contacts. Compared with the conventional methods based on forward dynamics, the proposed method reduces the computational cost of the dynamics and their sensitivities by utilizing the recursive Newton-Euler algorithm (RNEA) and its partial derivatives. In addition, it increases the sparsity of the Hessian of the Karush–Kuhn–Tucker conditions, which reduces the computational cost, e.g., of Riccati recursion. Numerical experiments show that the proposed method outperforms state-of-the-art implementations of differential dynamic programming based on forward dynamics in terms of computational time and numerical robustness.

Recent developments in agriculture have high-lighted the potential of as well as the need for the use of robotics. Various processes in this field can benefit from the proper use of state of the art technology [1], in terms of efficiency as well as quality. One of these areas is the harvesting of ripe fruit.In order to be able to automate this process, a robotic harvester needs to be aware of the full poses of the crop/fruit to be collected in order to perform proper path- and collision-planning. The current state of the art mainly considers problems of detection and segmentation of fruit with localisation limited to the 3D position only. The reliable and real-time estimation of the respective orientations remains a mostly unaddressed problem.In this paper, we present a compact and efficient network architecture for estimating the orientation of soft fruit such as strawberries from colour and, optionally, depth images. The proposed system can be automatically trained in a realistic simulation environment. We evaluate the system’s performance on simulated datasets and validate its operation on publicly available images of strawberries to demonstrate its practical use. Depending on the amount of training data used, coverage of state space, as well as the availability of RGB-D or RGB data only, mean errors of as low as 11° could be achieved.

Deep learning has been used to demonstrate end-to-end neural network learning for autonomous vehicle control from raw sensory input. While LiDAR sensors provide reliably accurate information, existing end-to-end driving solutions are mainly based on cameras since processing 3D data requires a large memory footprint and computation cost. On the other hand, increasing the robustness of these systems is also critical; however, even estimating the model’s uncertainty is very challenging due to the cost of sampling-based methods. In this paper, we present an efficient and robust LiDAR-based end-to-end navigation framework. We first introduce Fast-LiDARNet that is based on sparse convolution kernel optimization and hardware-aware model design. We then propose Hybrid Evidential Fusion that directly estimates the uncertainty of the prediction from only a single forward pass and then fuses the control predictions intelligently. We evaluate our system on a full-scale vehicle and demonstrate lane-stable as well as navigation capabilities. In the presence of out-of-distribution events (e.g., sensor failures), our system significantly improves robustness and reduces the number of takeovers in the real world.

A practical approach to robot reinforcement learning is to first collect a large batch of real or simulated robot interaction data, using some data collection policy, and then learn from this data to perform various tasks, using offline learning algorithms. Previous work focused on manually designing the data collection policy, and on tasks where suitable policies can easily be designed, such as random picking policies for collecting data about object grasping. For more complex tasks, however, it may be difficult to find a data collection policy that explores the environment effectively, and produces data that is diverse enough for the downstream task. In this work, we propose that data collection policies should actively explore the environment to collect diverse data. In particular, we develop a simple-yet-effective goal-conditioned reinforcement-learning method that actively focuses data collection on novel observations, thereby collecting a diverse data-set. We evaluate our method on simulated robot manipulation tasks with visual inputs and show that the improved diversity of active data collection leads to significant improvements in the downstream learning tasks.

Convolution has been used for fast computation of reachability maps, but it has high computational costs when performing SE(3) convolution operations for general joint arrangements in industrial robots and 3D workspace. Its application is also limited to planar robots, 2D workspace, or robots with special spatial arrangements for joints. In this paper, we find that the SE(3) convolution can be decomposed into a set of SE(2) convolutions, which significantly reduces the computational complexity when computing the reachability map of high-DOF robotic manipulators in the 3D workspace. We also leverage GPU parallel computing and Fast Fourier transform to further accelerate the computation procedure. We demonstrate the time efficiency and quality of our approach using a set of numerical experiments for constructing reachability maps and also present a multi-robot plant phenotyping system that uses the computed reachability map for efficient viewpoint selection and path planning.

We propose an efficient method to estimate the relative pose of a multi-camera system from a minimum of two affine correspondences (ACs). Our solution is novel as it computes the 6DOF relative pose by utilizing a first-order rotation approximation. We directly derive a single polynomial based on the constraint between ACs and the generalized camera model. Then a closed-form solution is found analytically and it produces an accurate relative pose estimation efficiently. Benefiting from the low number of exploited correspondences and the speed of the solver, it speeds up robust estimators, e.g. RANSAC, significantly. The proposed method is evaluated both on synthetic data and real-world image sequences from the KITTI benchmark. It is shown that the proposed solver is superior to the state-of-the-art algorithms in terms of accuracy.

It has been recently demonstrated that Temporal Convolution Networks (TCNs) provide state-of-the-art results in many problem domains where the input data is a time-series. TCNs typically incorporate information from a long history of inputs (the receptive field) into a single output using many convolution layers. Real-time inference using a trained TCN can be challenging on devices with limited compute and memory, especially if the receptive field is large. This paper introduces the RT-TCN algorithm that reuses the output of prior convolution operations to minimize the computational requirements and persistent memory footprint of a TCN during real-time inference. We also show that when a TCN is trained using time slices of the input time-series, it can be executed in realtime continually using RT-TCN. In addition, we provide TCN architecture guidelines that ensure that real-time inference can be performed within memory and computational constraints.

Neural Networks (NNs) can provide major empirical performance improvements for robotic systems, but they also introduce challenges in formally analyzing those systems’ safety properties. In particular, this work focuses on estimating the forward reachable set of closed-loop systems with NN controllers. Recent work provides bounds on these reachable sets, yet the computationally efficient approaches provide overly conservative bounds (thus cannot be used to verify useful properties), whereas tighter methods are too intensive for online computation. This work bridges the gap by formulating a convex optimization problem for reachability analysis for closed-loop systems with NN controllers. While the solutions are less tight than prior semidefinite program-based methods, they are substantially faster to compute, and some of the available computation time can be used to refine the bounds through input set partitioning, which more than overcomes the tightness gap. The proposed framework further considers systems with measurement and process noise, thus being applicable to realistic systems with uncertainty. Finally, numerical comparisons show that our approach based on linear programming and partitioning can give 10× reduction in conservatism in \frac{1}{2}\frac{1}{2} of the computation time compared to the state-of-the-art, and the ability to handle various sources of uncertainty is highlighted on a quadrotor model.

Analyzing scenes thoroughly is crucial for mobile robots acting in different environments. Semantic segmentation can enhance various subsequent tasks, such as (semantically assisted) person perception, (semantic) free space detection, (semantic) mapping, and (semantic) navigation. In this paper, we propose an efficient and robust RGB-D segmentation approach that can be optimized to a high degree using NVIDIA TensorRT and, thus, is well suited as a common initial processing step in a complex system for scene analysis on mobile robots. We show that RGB-D segmentation is superior to processing RGB images solely and that it can still be performed in real time if the network architecture is carefully designed. We evaluate our proposed Efficient Scene Analysis Network (ESANet) on the common indoor datasets NYUv2 and SUNRGB-D and show that we reach state-of-the-art performance while enabling faster inference. Furthermore, our evaluation on the outdoor dataset Cityscapes shows that our approach is suitable for other areas of application as well. Finally, instead of presenting benchmark results only, we also show qualitative results in one of our indoor application scenarios.

In this paper, we present an efficient online calibration system for longitudinal vehicle dynamics of driverless cars. Instead of modeling vehicle’s longitudinal dynamical system analytically, we employ a data-driven method to generate an "end-to-end" numerical model with a look-up table which saves vehicle’s velocity, control command, and acceleration. This reference table should be calibrated to account for variations of vehicle’s hardware status over time. To reduce the expensive labor in calibration process, we propose an effective algorithm to update this reference look-up table with a Gaussian model approach. We introduce a 2-D Gaussian distribution to model acceleration error between interpolated one from look-up table and actual one from vehicle sensors. We estimate model’s standard deviations with a "three-sigma rule" heuristic and calculate its height with a backtracking method such that monotonicity constraint between acceleration and control command is strictly satisfied in the updated table. The effectiveness of our proposed system is verified in realworld road tests with Lincoln MKZ.

In this paper, we design a versatile multi-sensor aided inertial navigation system (MINS) that can efficiently fuse multi-modal measurements of IMU, camera, wheel encoder, GPS, and 3D LiDAR along with online spatiotemporal sensor calibration. Building upon our prior work [1] –[3], in this work we primarily focus on efficient LiDAR integration in a sliding-window filtering fashion. As each 3D LiDAR scan contains a large volume of 3D points which poses great challenges for real-time performance, we advocate using plane patches, which contain the environmental structural information, extracted from the sparse LiDAR point cloud to update/calibrate the system efficiently. The proposed LiDAR plane patch processing algorithm (including extraction, data association, and update) is shown to be efficient and consistent. Both Extensive Monte-Carlo simulations and real-world datasets with large-scale urban driving scenarios have been used to verify the accuracy and consistency of the proposed MINS algorithm.

We present a novel hierarchical POMDP framework to solve an object search and delivery task where the agent is given a prior belief about the possible item locations. Solving POMDPs is computationally demanding and, as such, applications have typically been limited to small environments. The proposed hierarchical POMDP framework performs reasoning on multiple spatial scales in order to reduce computation time. The problem is first solved in the top layer of the hierarchy with a coarsely discretized state space. Its solution is refined in the lower layers with increasing resolution. Three different methods for propagating information down the spatial hierarchy are discussed and validated in simulation. We show that a two-layer multi-scale POMDP decreases computation time by an order of magnitude allowing for real-time applications while maintaining high solution quality. For large problems that require three layers to reach the desired resolution, computation time speedups by two orders of magnitude are achieved.

Modern agriculture relies on accurate and timely data. Currently, most of this data is gathered using remote sensing, which uses a combination of satellite and aerial imagery. However, ground robots are needed to fill in the gaps for finer ground-level data and the execution of physical tasks such as sample collection. The scales at which crops are produced preclude the inspection of each and every plant, thus requiring the selection of a smaller number of inspection targets. In this paper, we solve this multi-robot inspection problem using a novel task allocation algorithm. The algorithm derives its utility function from a model based on Gaussian process machine learning with a kernel that is learned from previous data. The algorithm also considers the physical limitations of moving within crop rows by dividing the plot into geodesic Voronoi regions based on robot locations. Simulation studies are performed to validate the method.

In many autonomous mapping tasks, the maps cannot be accurately constructed due to various reasons such as sparse, noisy, and partial sensor measurements. We propose a novel map prediction method built upon recent success of Low-Rank Matrix Completion. The proposed map prediction is able to achieve both map interpolation and extrapolation on raw poor-quality maps with missing or noisy observations. We validate with extensive simulated experiments that the approach can achieve real-time computation for large maps, and the performance is superior to state-of-the-art map prediction approach — Bayesian Hilbert Mapping in terms of mapping accuracy and computation time. Then we demonstrate that with the proposed real-time map prediction framework, the coverage convergence rate (per action step) for a set of representative coverage planning methods commonly used for environmental modeling and monitoring tasks can be significantly improved.

Robot path planning is difficult to solve due to the contradiction between the optimality of results and the complexity of algorithms, even in 2D environments. To find an optimal path, the algorithm needs to search all the state space, which costs many computation resources. To address this issue, we present a novel recurrent generative model (RGM), which generates efficient heuristic to reduce the search efforts of path planning algorithms. This RGM model adopts the framework of general generative adversarial networks (GAN), which consists of a novel generator that can generate heuristic by refining the outputs recurrently and two discriminators that check the connectivity and safety properties of heuristic. We test the proposed RGM module in various 2D environments to demonstrate its effectiveness and efficiency. The results show that, compared with a model without recurrence, the RGM successfully generates appropriate heuristic in both seen and new unseen maps with higher accuracy, demonstrating the good generalization ability of the RGM model. We also compare the rapidly-exploring random tree star (RRT*) with generated heuristic and the conventional RRT* in four different maps, showing that the generated heuristic can guide the algorithm to efficiently find both initial and optimal solutions in a faster and more efficient way.

Driven by the need for physically accurate and haptically convincing models of lunar and planetary regolith for model-mediated teleoperation in space, we present an approach to modelling regolith in an efficient yet realistic way. Model parameters are derived from physical characteristics of the regolith, to render regoliths with different density profiles, cohesion, internal friction and in different gravitational fields. Users could distinguish between changes in parameters– specifically friction and gravity field–and also gave qualitative feedback relevant to modelling regolith. We discuss the challenges in haptically rendering soils and the next steps.

The inverse dynamics of a robotic manipulator is instrumental in precise robot control and manipulation. However, acquiring such a model is challenging, not only due to unmodelled non-linearities such as joint friction, but also from a machine learning perspective (e.g., input space dimension, amount of data needed). The accuracy of such models, regardless of the learning techniques, relies on proper excitation and exploration of the robot’s configuration space, in order to collect a rich dataset. This study aims to provide rich data in learning the inverse dynamics of a serial robotic manipulator using supervised machine learning techniques. We propose a method, called Max-Information Configuration Exploration (MICE), to incrementally explore and generate information-rich data via computing parameters of a trajectory set. We also introduce a new set of excitation trajectories that explores robot’s configuration through imposed stable limit cycles in robot joints’ phase space while satisfying feasibility constraints and physical bounds. We benchmark MICE against state-of-the-art in terms of data quality and learning accuracy. The proposed methodology for data collection, model learning, and evaluation, is validated with a KUKA IIWA14 robotic arm where the results prove significant improvement over traditional approaches.