ICRA 2021 论文列表

Self-supervised goal proposal and reaching is a key component for exploration and efficient policy learning algorithms. Such a self-supervised approach without access to any oracle goal sampling distribution requires deep exploration and commitment so that long horizon plans can be efficiently discovered. In this paper, we propose an exploration framework, which learns a dynamics-aware manifold of reachable states. For a goal, our proposed method deterministically visits a state at the current frontier of reachable states (commit- ment/reaching) and then stochastically explores to reach the goal (exploration). This allocates exploration budget near the frontier of the reachable region instead of its interior. We target the challenging problem of policy learning from initial and goal states specified as images, and do not assume any access to the underlying ground-truth states of the robot and the environment. To keep track of reachable latent states, we propose a distance-conditioned reachability network that is trained to infer whether one state is reachable from another within the specified latent space distance. Given an initial state, we obtain a frontier of reachable states from that state. By incorporating a curriculum for sampling easier goals (closer to the start state) before more difficult goals, we demonstrate that the proposed self-supervised exploration algorithm, superior performance compared to existing baselines on a set of challenging robotic environments. https://sites.google.com/view/leaf-exploration

This paper investigates a hybrid solution which combines deep reinforcement learning (RL) and classical trajectory planning for the "following in front" application. Here, an autonomous robot aims to stay ahead of a person as the person freely walks around. Following in front is a challenging problem as the user’s intended trajectory is unknown and needs to be estimated, explicitly or implicitly, by the robot. In addition, the robot needs to find a feasible way to safely navigate ahead of human trajectory. Our deep RL module makes decisions at a high level by implicitly estimates the human trajectory and produces short-term navigational goals to guide the robot. These goals are used by a trajectory planner, which is responsible for low-level execution, to smoothly navigate the robot to the short-term goals, and eventually in front of the user. We employ curriculum learning in the deep RL module to efficiently achieve a high return. Our system outperforms the state-of-the-art in following ahead and is more reliable compared to end-to-end alternatives in both the simulation and real world experiments. In contrast to a pure deep RL approach, we demonstrate zero-shot transfer of the trained policy from simulation to the real world.

The process of learning a manipulation task depends strongly on the action space used for exploration: posed in the incorrect action space, solving a task with reinforcement learning can be drastically inefficient. Additionally, similar tasks or instances of the same task family impose latent manifold constraints on the most effective action space: the task family can be best solved with actions in a manifold of the entire action space of the robot. Combining these insights we present LASER, a method to learn latent action spaces for efficient reinforcement learning. LASER factorizes the learning problem into two sub-problems, namely action space learning and policy learning in the new action space. It leverages data from similar manipulation task instances, either from an offline expert or online during policy learning, and learns from these trajectories a mapping from the original to a latent action space. LASER is trained as a variational encoder-decoder model to map raw actions into a disentangled latent action space while maintaining action reconstruction and latent space dynamic consistency. We evaluate LASER on two contact-rich robotic tasks in simulation, and analyze the benefit of policy learning in the generated latent action space. We show improved sample efficiency compared to the original action space from better alignment of the action space to the task space, as we observe with visualizations of the learned action space manifold. Additional details: pair.toronto.edu/laser

Underwater image enhancement is an important low-level computer vision task for autonomous underwater vehicles and remotely operated vehicles to explore and understand the underwater environments. Recently, deep convolutional neural networks (CNNs) have been successfully used in many computer vision problems, and so does underwater image enhancement. There are many deep-learning-based methods with impressive performance for underwater image enhancement, but their memory and model parameter costs are hindrances in practical application. To address this issue, we propose a lightweight adaptive feature fusion network (LAFFNet). The model is the encoder-decoder model with multiple adaptive feature fusion (AAF) modules. AAF subsumes multiple branches with different kernel sizes to generate multi-scale feature maps. Furthermore, channel attention is used to merge these feature maps adaptively. Our method reduces the number of parameters from 2.5M to 0.15M (around 94% reduction) but outperforms state-of-the-art algorithms by extensive experiments. Furthermore, we demonstrate our LAFFNet effectively improves high-level vision tasks like salience object detection and single image depth estimation.

Koopman-based learning methods can potentially be practical and powerful tools for dynamical robotic systems. However, common methods to construct Koopman representations seek to learn lifted linear models that cannot capture nonlinear actuation effects inherent in many robotic systems. This paper presents a learning and control methodology that is a first step towards overcoming this limitation. Using the Koopman canonical transform, control-affine dynamics can be expressed by a lifted bilinear model. The learned model is used for nonlinear model predictive control (NMPC) design where the bilinear structure can be exploited to improve computational efficiency. The benefits for control-affine dynamics compared to existing Koopman-based methods are highlighted through an example of a simulated planar quadrotor. Prediction error is greatly reduced and closed loop performance similar to NMPC with full model knowledge is achieved.

Continuum manipulators have shown a wide range of applications due to their inherent compliance and dexterity. At present, the cross-sectional dimension of these manipulators is often kept constant, which facilitates the design, fabrication and modeling processes. The famous piecewise-constant-curvature (PCC) assumption is widely used in the kinematics and motion control. By contrast, most natural creatures that have continuum structures usually exhibit variable cross sections, such as the elephant trunk and the octopus tentacle. Inspired by these creatures, variable cross-section continuum manipulators have been proposed in literature. For variable cross-section, the PCC assumption is no longer valid. In fact, the manipulator’s kinematics and statics are coupled, which should be considered and solved together. This paper establishes the general kinetostatic model of variable cross-section continuum manipulators, and performs experimental validations under various actuating and loading situations. Results showed that the proposed model has satisfactory accuracy in predicting the manipulator profile under various actuating and loading situations. The average root-mean-square error of the proposed model was 1.67 mm, and even the maximum error did not exceed 2.6 mm, according to all the 51 groups of tests on three 100 mm long prototypes. Furthermore, results also indicated that Type-1 manipulator (only varying the spacer dimension) has larger deflections and tends to be "softer" than Type-2 manipulator (only varying the backbone cross-sectional area). Results of this paper could be helpful for the design and analysis of variable cross-section continuum manipulators.

Despite significant progress in the realm of upper-limb prosthetic design, end-users still abandon modern myoelectric prostheses, with haptic feedback listed as a primary need. The passive scheme of cable-driven body-powered prostheses provides kinesthetic sensory information to the user but can also lead to discomfort and fatigue due to the large loads applied to the body during operation. In order to investigate the role of this kinesthetic feedback on grasp force control, we design a body-powered prosthesis emulator capable of varying the amount of displayed force feedback along a continuous scale. Using this experimental test bed, we collect data from 9 participants while they perform a grasp and lift task. Analysis shows that, with increasing amounts of force feedback, people produce lower and steadier grasp forces but also become more prone to dropping held objects. These results suggest that the use of moderate amounts of feedback provides significant grasp performance benefits while also mitigating some of the shortcomings of cable-driven prostheses. These findings support the continued study of the incorporation of kinesthetic feedback into novel prosthetic designs.

Flexible surgical instruments can flexibly adjust their posture with a high degree of freedom, which makes them highly suitable for performing surgical tasks in narrow workspaces. However, redundant degrees of freedom increase their kinematic difficulty, which may cause redundant solutions, complex calculations, and low speeds. In this paper, a flexible surgical instrument is presented. The structural characteristics of this flexible instrument were explored in terms of force balance, it was concluded that the instrument had a constant curvature during bending. Based on this, the kinematics and inverse kinematics were solved via the geometric and Newton iteration methods, respectively. Our experiments showed that the proposed method for solving flexible instrument kinematics had high precision, a unique solution, and high speed; the instrument can be well controlled to perform refined operations. The proposed geometric method for solving the flexible instrument kinematics avoided the calculation of the Jacobian matrix, making it fast and capable of meeting the master-slave control requirement for real-time surgery. Furthermore, the proposed kinematics solution method is not limited by the mechanical structure, so it can be used for flexible instruments owning to its constant curvature bending.

Motion planning for Cable-Driven Parallel Robots (CDPRs) is a challenging task due to various restrictions on cable tensions, collisions and obstacle avoidance. The presented work aims at proposing an optimal path planning strategy in order to both maximize the wrench capability and the dexterity of the robot in a cluttered environment. First, an asymptoticallyoptimal path finding method based on a variant of rapidly exploring random trees (RRT) is implemented along with the GilbertJohnsonKeerthi (GJK) algorithm to account for the collision detections. Then, a goal biased Artificial Field Guide (AFG) is employed to reduce convergence time and ensure directional exploration. Finally, a post-processing algorithm is added to get a short and smooth resultant path by fitting appropriate splines. The proposed path planning strategy is analyzed and demonstrated on a simulated and experimental setup of a six-DOF spatial CDPR.

We present the first fully distributed multi-robot system for dense metric-semantic Simultaneous Localization and Mapping (SLAM). Our system, dubbed Kimera-Multi, is implemented by a team of robots equipped with visual-inertial sensors, and builds a 3D mesh model of the environment in real-time, where each face of the mesh is annotated with a semantic label (e.g., building, road, objects). In Kimera-Multi, each robot builds a local trajectory estimate and a local mesh using Kimera [1]. Then, when two robots are within communication range, they initiate a distributed place recognition and robust pose graph optimization protocol with a novel incremental maximum clique outlier rejection; the protocol allows the robots to improve their local trajectory estimates by leveraging inter-robot loop closures. Finally, each robot uses its improved trajectory estimate to correct the local mesh using mesh deformation techniques. We demonstrate Kimera-Multi in photo-realistic simulations and real data. Kimera-Multi (i) is able to build accurate 3D metric-semantic meshes, (ii) is robust to incorrect loop closures while requiring less computation than state-of-the-art distributed SLAM backends, and (iii) is efficient, both in terms of computation at each robot as well as communication bandwidth.

Mapping three-dimensional (3-D) dynamic environments is essential for aerial robots but challenging to consider the increased dimensions in both space and time compared to 2-D static mapping. This paper presents a kernel-based 3-D dynamic occupancy mapping algorithm, K3DOM, that distinguishes between static and dynamic objects while estimating the velocities of dynamic cells via particle tracking. The proposed algorithm brings the benefits of kernel inference such as its simple computation, consideration of spatial correlation, and natural measure of uncertainty to the domain of dynamic mapping. We formulate the dynamic occupancy mapping problem in a Bayesian framework and represent the map through Dirichlet distribution to update posteriors in a recursive way with intuitive heuristics. The proposed algorithm demonstrates its promising performance compared to baseline in diverse scenarios simulated in ROS environments.

Feature-based lidar odometry methods have attracted increasing attention due to their low computational cost. However, theoretically analysis of the effect of extracted features on pose estimation is still lacked. In this paper, we propose a method of key-feature selection for lightweight lidar inertial odometry, KFS-LIO, to further enhance the real-time performance by selecting the most effective subset of lidar feature constraints. Aiming at explaining the correlation between the feature distribution and state errors, a quantitative evaluation method of lidar constraints is introduced. In addition, to avoid recalculating the reprojection matrices in de-skewing step, we use the intermediate variables in IMU preintegration to compensate for lidar motion distortion. The experimental results demonstrate that KFS-LIO can reduce half of the LOAM features and provide comparable accuracy with the state-of-the-art odometry.

In this paper, we propose a new video object detection (VoD) method, referred to as temporal feature aggregation and motion-aware VoD (TM-VoD), that produces a joint representation of temporal image sequences and object motion. The TM-VoD generates strong spatiotemporal features for VOD by temporally redundant information in an image sequence and the motion context. These are produced at the feature level in the region proposal stage and at the instance level in the refinement stage. In the region proposal stage, visual features are temporally fused with appropriate weights at the pixel level via gated attention model. Furthermore, pixel level motion features are obtained by capturing the changes between adjacent visual feature maps. In the refinement stage, the visual features are aligned and aggregated at the instance level. We propose a novel feature alignment method, which uses the initial region proposals as anchors to predict the box coordinates for all video frames. Moreover, the instance level motion features are obtained by applying the region of interest (RoI) pooling to the pixel level motion features and by encoding the sequential changes in the box coordinates. Finally, all these instance level features are concatenated to produce a joint representation of the objects. Experiments on the ImageNet VID dataset demonstrate that the proposed method significantly outperforms existing VoDs and achieves performance comparable with that of state-of-the-art VoDs.

Object detection and data association are critical components in multi-object tracking (MOT) systems. Despite the fact that the two components are dependent on each other, prior works often design detection and data association modules separately which are trained with separate objectives. As a result, one cannot back-propagate the gradients and optimize the entire MOT system, which leads to sub-optimal performance. To address this issue, recent works simultaneously optimize detection and data association modules under a joint MOT framework, which has shown improved performance in both modules. In this work, we propose a new instance of joint MOT approach based on Graph Neural Networks (GNNs). The key idea is that GNNs can model relations between variablesized objects in both the spatial and temporal domains, which is essential for learning discriminative features for detection and data association. Through extensive experiments on the MOT15/16/17/20 datasets, we demonstrate the effectiveness of our GNN-based joint MOT approach and show state-of-the-art performance for both detection and MOT tasks.

Single shot approaches have demonstrated tremendous success on various computer vision tasks. Finding good parameterizations for 6D object pose estimation remains an open challenge. In this work, we propose different novel parameterizations for the output of the neural network for single shot 6D object pose estimation. Our learning-based approach achieves state-of-the-art performance on two public benchmark datasets. Furthermore, we demonstrate that the pose estimates can be used for real-world robotic grasping tasks without additional ICP refinement.

Unmanned Aerial Vehicle (UAV) flight paths have been shown to communicate meaning to human observers, similar to human gestural communication. This paper presents the results of a UAV gesture perception study designed to assess how observer viewpoint perspective may impact how humans perceive the shape of UAV gestural motion. Robot gesture designers have demonstrated that robots can indeed communicate meaning through gesture; however, many of these results are limited to an idealized range of viewer perspectives and do not consider how the perception of a robot gesture may suffer from obfuscation or self-occlusion from some viewpoints. This paper presents the results of three online user-studies that examine participants' ability to accurately perceive the intended shape of two-dimensional UAV gestures from varying viewer perspectives. We used a logistic regression model to characterize participant gesture classification accuracy, demonstrating that viewer perspective does impact how participants perceive the shape of UAV gestures. Our results yielded a viewpoint angle threshold from beyond which participants were able to assess the intended shape of a gesture's motion with 90% accuracy. We also introduce a perceptibility score to capture user confidence, time to decision, and accuracy in labeling and to understand how differences in flight paths impact perception across viewpoints. These findings will enable UAV gesture systems that, with a high degree of confidence, ensure gesture motions can be accurately perceived by human observers.

Robot interfaces often only use the visual channel. Inspired by Wickens’ Multiple Resource Theory, we investigated if the addition of audio elements would reduce cognitive workload and improve performance. Specifically, we designed a search and threat-defusal task (primary) with a memory test task (secondary). Eleven participants – predominantly first responders – were recruited to control a robot to clear all threats in a combination of four conditions of primary and secondary tasks in visual and auditory channels. While we did not find any statistically significant differences in performance or workload across subjects, making it questionable that Multiple Resource Theory could shorten longer-term task completion time and reduce workload. Our results suggest that considering individual differences for splitting interface modalities across multiple channels requires further investigation.

Benchmarks of state-of-the-art rigid-body dynamics libraries report better performance solving the inverse dynamics problem than the forward alternative. Those benchmarks encouraged us to question whether that computational advantage would translate to direct transcription, where calculating rigid-body dynamics and their derivatives accounts for a significant share of computation time. In this work, we implement an optimization framework where both approaches for enforcing the system dynamics are available. We evaluate the performance of each approach for systems of varying complexity, for domains with rigid contacts. Our tests reveal that formulations using inverse dynamics converge faster, require less iterations, and are more robust to coarse problem discretization. These results indicate that inverse dynamics should be preferred to enforce the nonlinear system dynamics in simultaneous methods, such as direct transcription.

We present a trajectory planning and control architecture for bipedal locomotion at a variety of speeds on a highly underactuated and compliant bipedal robot. A library of compliant walking trajectories are planned offline, and stored as compact arrays of polynomial coefficients for tracking online. The control implementation uses a floating-base inverse dynamics controller which generates dynamically consistent feedforward torques to realize walking using information obtained from the trajectory optimization. The effectiveness of the controller is demonstrated in simulation and on hardware for walking both indoors on flat terrain and over unplanned disturbances outdoors. Additionally, both the controller and optimization source code are made available on GitHub.