ICRA 2021 论文列表

This paper proposes a systematic solution that uses an unmanned aerial vehicle (UAV) to aggressively and safely track an agile target. It properly handles the challenging situations where the intent of the target and the dense environments are unknown. Our work is divided into two parts: target motion prediction and tracking trajectory planning. The target motion prediction method utilizes target observations to reliably predict its future motion. The tracking trajectory planner follows the hierarchical workflow. A target informed kinody-namic searching method is adopted as the front-end, which heuristically searches for a safe tracking trajectory. The back- end optimizer then refines it into a spatial-temporal optimal trajectory. The proposed solution is integrated into an onboard quadrotor system. We fully test the system in challenging real-world tracking missions. Moreover, benchmark comparisons validate that the proposed method surpasses the cutting-edge methods on time efficiency and tracking effectiveness.

Deep learning-based object pose estimators are often unreliable and overconfident especially when the input image is outside the training domain, for instance, with sim2real transfer. Efficient and robust uncertainty quantification (UQ) in pose estimators is critically needed in many robotic tasks. In this work, we propose a simple, efficient, and plug-and-play UQ method for 6-DoF object pose estimation. We ensemble 2–3 pre-trained models with different neural network architectures and/or training data sources, and compute their average pair-wise disagreement against one another to obtain the uncertainty quantification. We propose four disagreement metrics, including a learned metric, and show that the average distance (ADD) is the best learning-free metric and it is only slightly worse than the learned metric, which requires labeled target data. Our method has several advantages compared to the prior art: 1) our method does not require any modification of the training process or the model inputs; and 2) it needs only one forward pass for each model. We evaluate the proposed UQ method on three tasks where our uncertainty quantification yields much stronger correlations with pose estimation errors than the baselines. Moreover, in a real robot grasping task, our method increases the grasping success rate from 35% to 90%. Video and code are available at https://sites.google.com/view/fastuq.

In this paper, the experimental swing-up of a spherical pendulum mounted on a collaborative robot is presented. The complete mechanical system consists of nine degrees of freedom (DoFs). The primary focus of this work is the design of a fast trajectory planning for the swing-up by systematically incorporating the kinematic and dynamics constraints. The proposed algorithm consists of two steps: First, an offline trajectory optimization is used to build a database of swing-up trajectories, with an average computing time of 10 s for one trajectory. Second, a fast trajectory replanner based on a constrained quadratic program is described, which computes the swing-up trajectory for an arbitrary initial configuration of the system with an average computing time of 0.2 s. Simulations and experimental results demonstrate the swing-up of the spherical pendulum using a discrete time-variant linear quadratic regulator as a feedback controller.

In this work, we present a new exploration algorithm for Micro Aerial Vehicles (MAVs). The planner uses a combination of Next-Best-View (NBV) sampling and frontier-based approaches to reduce the impact of finding unexplored areas in large scenarios. For each sampled point, the yaw angle is optimized to maximize the potential gain for mapping. The gain is expressed as a ratio between the exploration objective and the time it would take to reach the pose, thus, balancing the nearby exploration and global coverage. We reduce the gain computation bottleneck in the sampling strategy by managing a dual-map with different resolutions. The planner maintains a history graph, with nodes that indicate regions of interest for map expansion. We demonstrate the abilities of the proposed algorithm with simulations and real-world experiments. The results outperform state-of-the-art methods in exploration time and computational cost.

In this paper, we proposed a novel path replanning algorithm on arbitrary graphs. To avoid computationally heavy preprocessing and to reduce required memory to store the expanded vertices of the previous search, we defined the feature vertices, which are extracted from the previous path by a simple algorithm to compare the costs between adjacent vertices along the path once. Proper additional heuristic functions are designed for these feature vertices to work as local attractors guiding the search toward the previous path’s neighbors. To avoid unnecessary expansions and speed up the search, these additional heuristic functions are properly managed to stop intriguing or guiding search toward the feature vertices. The proposed algorithm of Fast Replanning Multi-Heuristic A* is a variation of Shared Multi-Heuristic A* while removing or deactivating the additional heuristic functions during the search. Fast Replanning Multi-Heuristic A* guarantees the bounded suboptimality while efficiently exploring the graph toward the goal vertex. The performance of the proposed algorithm was compared with weighted A* and D* lite by simulating numerous path replanning problems in maze-like maps.

Fast autonomous motion in cluttered and unknown environments, such as forests, is highly dependent on low-latency obstacle avoidance strategies. In this context, this paper presents a motion planning strategy that relies on lattices for the fast computation of local paths that both avoid obstacles and follow a vector field that encodes the global robot task. Lattices are constructed in the sensor space and represent a set of search trees that can be quickly pruned in function of the detected obstacles. The remaining lattice trees are used to optimize a vector field-dependent functional, thus generating the best free local path that tracks the field. To illustrate the proposed approach, we present simulation and real-world experiments of a planar robot moving in a cluttered, forest-like environment.

Object segmentation is a key component in the visual system of a robot that performs tasks like grasping and object manipulation, especially in presence of occlusions. Like many other computer vision tasks, the adoption of deep architectures has made available algorithms that perform this task with remarkable performance. However, adoption of such algorithms in robotics is hampered by the fact that training requires large amount of computing time and it cannot be performed on-line.In this work, we propose a novel architecture for object segmentation, that overcomes this problem and provides comparable performance in a fraction of the time required by the state-of-the-art methods. Our approach is based on a pre-trained Mask R-CNN, in which various layers have been replaced with a set of classifiers and regressors that are retrained for a new task. We employ an efficient Kernel-based method that allows for fast training on large scale problems. Our approach is validated on the YCB-Video dataset which is widely adopted in the computer vision and robotics community, demonstrating that we can achieve and even surpass performance of the state-of-the-art, with a significant reduction (~6×) of the training time.The code to reproduce the experiments is publicly available on GitHub1.

Multi-robot systems are uniquely well-suited to performing complex tasks such as patrolling and tracking, information gathering, and pick-up and delivery problems, offering significantly higher performance than single-robot systems. A fundamental building block in most multi-robot systems is task allocation: assigning robots to tasks (e.g., patrolling an area, or servicing a transportation request) as they appear based on the robots’ states to maximize reward. In many practical situations, the allocation must account for heterogeneous capabilities (e.g., availability of appropriate sensors or actuators) to ensure the feasibility of execution, and to promote a higher reward, over a long time horizon. To this end, we present the FlowDec algorithm for efficient heterogeneous task-allocation, and show that it achieves an approximation factor of at least 1/2 of the optimal reward. Our approach decomposes the heterogeneous problem into several homogeneous subproblems that can be solved efficiently using min-cost flow. Through simulation experiments, we show that our algorithm is faster by several orders of magnitude than a MILP approach.

This paper presents a Dynamic Vision Sensor (DVS) based system for reasoning about high-speed motion. As a representative scenario we consider a robot at rest, reacting to a small, fast approaching object at speeds higher than 15 m/s. Since conventional image sensors at typical frame rates observe such an object for only a few frames, estimating the underlying motion presents a considerable challenge for standard computer vision systems and algorithms. We present a method motivated by how animals such as insects solve this problem with their relatively simple vision systems.Our solution takes the event stream from a DVS and first encodes the temporal events with a set of causal exponential filters across multiple time scales. We couple these filters with a Convolutional Neural Network (CNN) to efficiently extract relevant spatiotemporal features. The combined network learns to output both the expected time to collision of the object, as well as the predicted collision point on a discretized polar grid. These critical estimates are computed with minimal delay by the network in order to react appropriately to the incoming object. We highlight our system’s results with a toy dart moving at 23.4 m/s with a 24.73° error in θ, 18.4 mm average discretized radius prediction error, and 25.03% median time to collision prediction error.

This paper aims to present a drone swarm light show design platform to support STEAM (science, technology, engineering, art and mathematics) education for K-12 children. With this platform, children can use this platform to design a drone swarm light show easily. To this end, the architecture of this platform contents three layers: UI layer, command layer, and physical layer. The UI layer has an easy-to-use interface for children. Children can feed parameters about the light show by clicking buttons and dragging sliders of four tracks. All actions designed for the swarm in the UI layer will be generated automatically in the form of the drone’s desired trajectories through the command layer. The physical layer includes a router for communication and a drone swarm for the light show. Our experimental results demonstrate that this platform works efficiently and suits for being applied to real STEAM education.

Footstep planning is the dominating approach when it comes to controlling the walk of a humanoid robot, even though a footstep plan is expensive to compute. The most prominent proposals typically spend up to a few seconds of computation time and output a sequence of up to 30 steps all the way to the goal. This way, footstep planning is applicable only in static environments where nothing changes after a plan has been computed. Since uncontrolled environments present challenges such as unforeseen motion of other objects and unexpected disturbances to balance, fast replanning of a footstep plan while the robot is in motion is highly desirable. We present a new way of fast footstep planning - Aborting A* - which is able to guarantee a replanning rate of 50 Hz by aborting an A* search before completion. We make aborting possible by using a novel, obstacle-aware heuristic function that lays out rotate-translate-rotate motions along the shortest path to the goal, enabling us to stop the planning progress prematurely with a target-oriented solution at any time during the search, even after only a few nodes have been expanded. We show in our experiments that despite the bounded computation time, our planner computes good results and does not get stuck in local minima.

Autonomous agents interacting with the real world need to learn new concepts efficiently and reliably. This requires learning in a low-data regime, which is a highly challenging problem. We address this task by introducing a fast optimization-based meta-learning method for few-shot classification. It consists of an embedding network, providing a general representation of the image, and a base learner module. The latter learns a linear classifier during the inference through an unrolled optimization procedure. We design an inner learning objective composed of (i) a robust classification loss on the support set and (ii) an entropy loss, allowing transductive learning from unlabeled query samples. By employing an efficient initialization module and a Steepest Descent based optimization algorithm, our base learner predicts a powerful classifier within only a few iterations. Further, our strategy enables important aspects of the base learner objective to be learned during meta-training. To the best of our knowledge, this work is the first to integrate both induction and transduction into the base learner in an optimization-based meta-learning framework. We perform a comprehensive experimental analysis, demonstrating the speed and effectiveness of our approach on four few-shot classification datasets. The Code is available at https://github.com/4rdhendu/FIML.

Roboticists have increasingly sought to incorporate mechanical compliance into legged robots to realize a range of potential benefits, from improved agility to resilience in complex environments. A promising approach for building compliance into robot legs is to utilize the pneumatic artificial muscle, a pneumatic actuator with inherent compliance due to the compressibility of air. While previous work has explored the capabilities of pneumatic-muscle driven robots in highly dynamic tasks like jumping, there is a lack of trajectory planning strategies for such robots. In this paper, we detail our approach to planning vertical jumping trajectories for a planar two-legged robot driven by four pneumatic artificial muscles using on/off "burst inflation" control. The trajectory optimization problem is represented as a factor graph and solved with the GTSAM optimizer. A hybrid dynamics approach is used to handle foot-ground contacts. The average jump height error between simulation and experiment across multiple jumping trajectories of varying heights was 9.5 cm; the average RMS error between all four joints was 5.6 deg. This work provides a basis to plan more complex jumping and leaping trajectories for pneumatic muscle-driven robots.

This paper proposes a method of fabric structure defect detection based on tactile information. Different from traditional visual-based detection methods, the proposed method uses a tactile sensor to attain the information of fabric. The advantage of using tactile information is to avoid different irregular dyeing patterns and reduce the influence of ambient light. Therefore, the proposed method can be more concise and universal, which makes the defect detection system more robust. Experiments are conducted to verify the performance of the developed tactile method. The results demonstrate that the proposed method is much better than the visual method in the detection of structural defects. In addition, we propose a design of a tactile sensing device that can greatly improve the actual detection efficiency of the tactile method. It is showed that the proposed method is efficient enough to meet our requirements and it provides a possible solution for the application of the tactile method in actual fabric production.

Cervical spinal cord injury (SCI) can severely impact hand motor and sensory function, and accordingly, patients with SCI are often unable to complete basic everyday tasks without assistance. In recent years, there has been an increase in hand exoskeleton research due to their distinct advantages for improving rehabilitation. In this work, we present a voice-controlled, tendon-driven soft robotic hand rehabilitation exoskeleton for hand function improvement in patients with cervical SCI. A new fabrication process utilizing high consistency rubber silicone is used to construct a formfitting, durable, and customizable exoskeleton glove. Bioinspired tendon routing pathways embedded within the glove create 5 actively actuated degrees-of-freedom in the index finger, middle finger, and thumb for both extension and flexion. Tendon tension sensors are developed and integrated into the exoskeleton system. The force feedback from the sensors is used in the implementation of admittance control for the exoskeleton system to improve grasping motions. The exoskeleton was evaluated in a case study with a healthy participant through range-of-motion characterization, pinch force testing, admittance controller validation, and object manipulation.

This paper investigates reinforcement learning with constraints, which are indispensable in safety-critical environments. To drive the constraint violation to decrease monotonically, we take the constraints as Lyapunov functions and impose new linear constraints on the policy parameters’ updating dynamics. As a result, the original safety set can be forward-invariant. However, because the new guaranteed-feasible constraints are imposed on the updating dynamics instead of the original policy parameters, classic optimization algorithms are no longer applicable. To address this, we propose to learn a generic deep neural network (DNN)-based optimizer to optimize the objective while satisfying the linear constraints. The constraint-satisfaction is achieved via projection onto a polytope formulated by multiple linear inequality constraints, which can be solved analytically with our newly designed metric. To the best of our knowledge, this is the first DNN-based optimizer for constrained optimization with the forward invariance guarantee. We show that our optimizer trains a policy to decrease the constraint violation and maximize the cumulative reward monotonically. Results on numerical constrained optimization and obstacle-avoidance navigation validate the theoretical findings.

In this paper, we investigate the problem of weakly supervised 3D vehicle detection. Conventional methods for 3D object detection usually require vast amounts of manually labelled 3D data as supervision signals. However, annotating large datasets needs huge human efforts, especially for 3D area. To tackle this problem, we propose a frustum-aware geometric reasoning (FGR) method to detect vehicles in point clouds without any 3D annotations. Our method consists of two stages: coarse 3D segmentation and 3D bounding box estimation. For the first stage, a context-aware adaptive region growing algorithm is designed to segment objects based on 2D bounding boxes. Leveraging predicted segmentation masks, we develop an anti-noise approach to estimate 3D bounding boxes in the second stage. Finally 3D pseudo labels generated by our method are utilized to train a 3D detector. Independent of any 3D groundtruth, FGR reaches comparable performance with fully supervised methods on the KITTI dataset. The findings indicate that it is able to accurately detect objects in 3D space with only 2D bounding boxes and sparse point clouds.

This work presents a general deep learning framework for large-scale point clouds understanding without voxelizations, called FG-Conv, which achieves an accurate and real-time understanding of point clouds. Through our novel design combining feature level correlation mining and deformable convolutions based geometric aware modeling, the local feature relationships and geometric patterns can be captured. The attention mechanism is also adopted to enhance the global long-range feature correlations. Finally, the feature pyramid residual learning network is proposed to combine patterns at different resolutions in a memory-efficient way. Extensive experiments on real-world challenging datasets demonstrated that our approaches outperform state-of-the-art methods in terms of accuracy and efficiency. Weakly supervised transfer learning demonstrates the generalization capacity of our methods.

Deep learning has achieved remarkable success in object recognition tasks through the availability of large scale datasets like ImageNet. However, deep learning systems suffer from catastrophic forgetting when learning incrementally without replaying old data. For real-world applications, robots also need to incrementally learn new objects. Further, since robots have limited human assistance available, they must learn from only a few examples. However, very few object recognition datasets and benchmarks exist to test incremental learning capability for robotic vision. Further, there is no dataset or benchmark specifically designed for incremental object learning from a few examples. To fill this gap, we present a new dataset termed F-SIOL-310 (Few-Shot Incremental Object Learning) which is specifically captured for testing few-shot incremental object learning capability for robotic vision. We also provide benchmarks and evaluations of 8 incremental learning algorithms on F-SIOL-310 for future comparisons. Our results demonstrate that the few-shot incremental object learning problem for robotic vision is far from being solved.