IROS 2022 Papers

Modern autonomous vehicles (AVs) often rely on vision, LIDAR, and even radar-based simultaneous localization and mapping (SLAM) frameworks for precise localization and navigation. However, modern SLAM frameworks often lead to unacceptably high levels of drift (i.e., localization error) when AVs observe few visually distinct features or encounter occlusions due to dynamic obstacles. This paper argues that minimizing drift must be a key desiderata in AV motion planning, which requires an AV to take active control decisions to move towards feature-rich regions while also minimizing conventional control cost. To do so, we first introduce a novel data-driven perception module that observes LIDAR point clouds and estimates which features/regions an AV must navigate towards for drift minimization. Then, we introduce an interpretable model predictive controller (MPC) that moves an AV toward such feature-rich regions while avoiding visual occlusions and gracefully trading off drift and control cost. Our experiments on challenging, dynamic scenarios in the state-of-the-art CARLA simulator indicate our method reduces drift up to 76.76% compared to benchmark approaches.

The present paper proposes a novel reinforce-ment learning method with world models, DreamingV2, a collaborative extension of DreamerV2 and Dreaming. Dream- erV2 is a cutting-edge model-based reinforcement learning from pixels that uses discrete world models to represent latent states with categorical variables. Dreaming is also a form of reinforcement learning from pixels that attempts to avoid the auto encoding process in general world model training by involving a reconstruction-free contrastive learning objective. The proposed DreamingV2 is a novel approach of adopting both the discrete representation of DreamingV2 and the reconstruction-free objective of Dreaming. Compared to DreamerV2 and other recent model-based methods without reconstruction, DreamingV2 achieves the best scores on five simulated challenging 3D robot arm tasks. We believe that DreamingV2 will be a reliable solution for robot learning since its discrete representation is suitable to describe discontinuous environments, and the reconstruction-free fashion well manages complex vision observations.

Tracking position and orientation independently affords more agile maneuver for over-actuated multirotor Unmanned Aerial Vehicles (UAVs) while introducing undesired downwash effects; downwash flows generated by thrust generators may counteract others due to close proximity, which significantly threatens the stability of the platform. The complexity of modeling aerodynamic airflow challenges control algorithms from properly compensating for such a side effect. Leveraging the input redundancies in over-actuated UAVs, we tackle this issue with a novel control allocation framework that considers downwash effects and explores the entire allocation space for an optimal solution. This optimal solution avoids downwash effects while providing high thrust efficiency within the hardware constraints. To the best of our knowledge, ours is the first formal derivation to investigate the downwash effects on over-actuated UAVs. We verify our framework on different hardware configurations in both simulation and experiment.

Visual localization is an important task for many robotic and augmented reality applications. As localizing within large scale maps can be memory and computationally de-manding, cloud-based localization services are appealing for developers. However, such services raise important privacy concerns for both passive and active users. In particular, some sensitive information might be revealed by an attacker who intercepts data during the data-sharing process. Therefore, the sensitive data in the image should be concealed before it is shared. As a motivating case, we demonstrated the exposure generated by the common feature descriptor SIFT when attempting to recover private content. In this paper, we propose a pipeline to effectively conceal privacy-sensitive image regions from possible attacks to the transmission or localization services, by making use of learning-based image inpainting techniques while preserving, and even boosting, the localization performance. We tested our pipeline with two off-the-shelf localization services based on deep neural networks on the publicly available Oxford Robotcar dataset, showing that the localization performance on our generated private concealed images is on par with the non-private baseline. 1

Motion forecasting in highly interactive scenarios is a challenging problem in autonomous driving. In such scenarios, we need to accurately predict the joint behavior of interacting agents to ensure the safe and efficient navigation of autonomous vehicles. Recently, goal-conditioned methods have gained increasing attention due to their advantage in performance and their ability to capture the multimodality in trajec-tory distribution. In this work, we study the joint trajectory prediction problem with the goal-conditioned framework. In particular, we introduce a conditional-variational-autoencoder-based (CVAE) model to explicitly encode different interaction modes into the latent space. However, we discover that the vanilla model suffers from posterior collapse and cannot induce an informative latent space as desired. To address these issues, we propose a novel approach to avoid KL vanishing and induce an interpretable interactive latent space with pseudo labels. The proposed pseudo labels allow us to incorporate domain knowledge on interaction in a flexible manner. We motivate the proposed method using an illustrative toy example. In addition, we validate our framework on the Waymo Open Motion Dataset with both quantitative and qualitative evaluations.

The problem of image change detection via every-day indoor robot navigation is explored from a novel perspective of the self-attention technique. Detecting semantically non-distinctive and visually small changes remains a key challenge in the robotics community. Intuitively, these small non-distinctive changes may be better handled by the recent paradigm of the attention mechanism, which is the basic idea of this work. However, existing self-attention models require significant retraining cost per domain, so it is not directly applicable to robotics applications. We propose a new self-attention technique with an ability of unsupervised on-the-fly domain adaptation, which introduces an attention mask into the intermediate layer of an image change detection model, without modifying the input and output layers of the model. Experiments, in which an indoor robot aims to detect visually small changes in everyday navigation, demonstrate that our attention technique significantly boosts the state-of-the-art image change detection model. Our datset is available at https://github.com/KojiTakeda00/Small_object_change_detection

When cast into the Deep Reinforcement Learning framework, many robotics tasks require solving a long horizon and sparse reward problem, where learning algorithms struggle. In such context, Imitation Learning (IL) can be a powerful approach to bootstrap the learning process. However, most IL methods require several expert demonstrations which can be prohibitively difficult to acquire. Only a handful of IL algorithms have shown efficiency in the context of an extreme low expert data regime where a single expert demonstration is available. In this paper, we present a novel algorithm designed to imitate complex robotic tasks from the states of an expert trajectory. Based on a sequential inductive bias, our method divides the complex task into smaller skills. The skills are learned into a goal-conditioned policy that is able to solve each skill individually and chain skills to solve the entire task. We show that our method imitates a non-holonomic navigation task and scales to a complex simulated robotic manipulation task with very high sample efficiency.

We present the first distributed optimization al-gorithm with lazy communication for collaborative geometric estimation, the backbone of modern collaborative simultaneous localization and mapping (SLAM) and structure-from-motion (SfM) applications. Our method allows agents to cooperatively reconstruct a shared geometric model on a central server by fusing individual observations, but without the need to transmit potentially sensitive information about the agents themselves (such as their locations). Furthermore, to alleviate the burden of communication during iterative optimization, we design a set of communication triggering conditions that enable agents to selectively upload a targeted subset of local information that is useful to global optimization. Our approach thus achieves significant communication reduction with minimal impact on optimization performance. As our main theoretical contribution, we prove that our method converges to first-order critical points with a global sublinear convergence rate. Numerical evaluations on bundle adjustment problems from collaborative SLAM and SfM datasets show that our method performs competitively against existing distributed techniques, while achieving up to 78% total communication reduction.

To accomplish task efficiently in a multiple robots system, a problem that has to be addressed is Simultaneous Localization and Mapping (SLAM). LiDAR (Light Detection and Ranging) has been used for many SLAM solutions due to its superb accuracy, but its performance degrades in featureless environments, like tunnels or long corridors. Centralized SLAM solves the problem with a cloud server, which requires a huge amount of computational resources and lacks robustness against central node failure. To address these issues, we present a distributed SLAM solution to estimate the trajectory of a group of robots using Ultra-WideBand (UWB) ranging and odometry measurements. The proposed approach distributes the processing among the robot team and significantly mitigates the computation concern emerged from the centralized SLAM. Our solution determines the relative pose (also known as loop closure) between two robots by minimizing the UWB ranging measurements taken at different positions when the robots are in close proximity. UWB provides a good distance measure in line-of-sight conditions, but retrieving a precise pose estimation remains a challenge, due to ranging noise and unpredictable path traveled by the robot. To deal with the suspicious loop closures, we use Pairwise Consistency Maximization (PCM) to examine the quality of loop closures and perform outlier rejections. The filtered loop closures are then fused with odometry in a distributed pose graph optimization (DPGO) module to recover the full trajectory of the robot team. Extensive experiments are conducted to validate the effectiveness of the proposed approach.

Reinforcement learning commonly suffers from slow convergence speed and requires thousands of episodes, which makes it hard to be applied for physical robotic applications. Little research has been studied for snake robot control using RL because of the additional difficulty of high redundancy of freedom. Existing methods either adopts an asynchronous A3C structure or a joint state representation. We propose a distributed coach-based deep learning method for snake robot control, which can greatly expedite the training speed with less episodes. The major contributions include: 1) a completely distributed graphical formulation; 2) an explicit stochastic density propagation rule for each robot link; 3) various interaction models with uncertainty estimation. The preliminary results of both simulation and real-world experiments have demonstrated the promising performance in comparison with state-of-the-art.

Estimating precise metric depth and scene reconstruction from monocular endoscopy is a fundamental task for surgical navigation in robotic surgery. However, traditional stereo matching adopts binocular images to perceive the depth information, which is difficult to transfer to the soft robotics-based surgical systems due to the use of monocular endoscopy. In this paper, we present a novel framework that combines robot kinematics and monocular endoscope images with deep unsupervised learning into a single network for metric depth estimation and then achieve 3D reconstruction of complex anatomy. Specifically, we first obtain the relative depth maps of surgical scenes by leveraging a brightness-aware monocular depth estimation method. Then, the corresponding endoscope poses are computed based on non-linear optimization of geo-metric and photometric reprojection residuals. Afterwards, we develop a Depth-driven Sliding Optimization (DDSO) algorithm to extract the scaling coefficient from kinematics and calculated poses offline. By coupling the metric scale and relative depth data, we form a robust ensemble that represents the metric and consistent depth. Next, we treat the ensemble as supervisory labels to train a metric depth estimation network for surgeries (i.e., MetricDepthS-Net) that distills the embeddings from the robot kinematics, endoscopic videos, and poses. With accurate metric depth estimation, we utilize a dense visual reconstruction method to recover the 3D structure of the whole surgical site. We have extensively evaluated the proposed framework on public SCARED and achieved comparable performance with stereo-based depth estimation methods. Our results demon-strate the feasibility of the proposed approach to recover the metric depth and 3D structure with monocular inputs.

In this paper, we propose a new sampling-based path planning approach, focusing on the challenges linked to autonomous exploration. Our method relies on the definition of a disk graph of free-space bubbles, from which we derive a biased sampling function that expands the graph towards known free space for maximal navigability and frontiers discovery. The proposed method demonstrates an exploratory behavior similar to Rapidly-exploring Random Trees, while retaining the connectivity and flexibility of a graph-based planner. We demonstrate the interest of our method by first comparing its path planning capabilities against state-of-the-art approaches, before discussing exploration-specific aspects, namely replanning capabilities and incremental construction of the graph. A simple frontiers-driven exploration controller derived from our planning method is also demonstrated using the Pioneer platform.

Equipping robots with the ability to infer human intent is a vital precondition for effective collaboration. Most computational approaches towards this objective derive a probability distribution of “intent” conditioned on the robot's perceived state. However, these approaches typically assume task-specific labels of human intent are known a priori. To overcome this constraint, we propose the Disentangled Sequence Clustering Variational Autoencoder (DiSCVAE), a clustering framework capable of learning such a distribution of intent in an unsupervised manner. The proposed framework leverages recent advances in unsupervised learning to disentangle latent representations of sequence data, separating time-varying local features from time-invariant global attributes. As a novel extension, the DiSCVAE also infers a discrete variable to form a latent mixture model and thus enable clustering over these global sequence concepts, e.g. high-level intentions. We evaluate the DiSCVAE on a real-world human-robot interaction dataset collected using a robotic wheelchair. Our findings reveal that the inferred discrete variable coincides with human intent, holding promise for collaborative settings, such as shared control.

We propose a novel approach for planning agents to compose abstract skills via observing and learning from historical interactions with the world. Our framework operates in a Markov state-space model via a set of actions under unknown pre-conditions. We formulate skills as high-level abstract policies that propose action plans based on the current state. Each policy learns new plans by observing the states' transitions while the agent interacts with the world. Such an approach automatically learns new plans to achieve specific intended effects, but the success of such plans is often dependent on the states in which they are applicable. Therefore, we formulate the evaluation of such plans as infinitely many multi-armed bandit problems, where we balance the allocation of resources on evaluating the success probability of existing arms and exploring new options. The result is a planner capable of automatically learning robust high-level skills under a noisy environment; such skills implicitly learn the action pre-condition without explicit knowledge. We show that this planning approach is experimentally very competitive in high-dimensional state space domains.

This paper describes the practical response- and performance-aware development of online speech enhancement for an augmented reality (AR) headset that helps a user understand conversations made in real noisy echoic environments (e.g., cocktail party). One may use a state-of-the-art blind source separation method called fast multichannel nonnegative matrix factorization (FastMNMF) that works well in various environments thanks to its unsupervised nature. Its heavy computational cost, however, prevents its application to real-time processing. In contrast, a supervised beamforming method that uses a deep neural network (DNN) for estimating spatial information of speech and noise readily fits real-time processing, but suffers from drastic performance degradation in mismatched conditions. Given such complementary characteristics, we propose a dual-process robust online speech enhancement method based on DNN-based beamforming with FastMNMF-guided adaptation. FastMNMF (back end) is performed in a mini-batch style and the noisy and enhanced speech pairs are used together with the original parallel training data for updating the direction-aware DNN (front end) with backpropagation at a computationally-allowable interval. This method is used with a blind dereverberation method called weighted prediction error (WPE) for transcribing the noisy reverberant speech of a speaker, which can be detected from video or selected by a user's hand gesture or eye gaze, in a streaming manner and spatially showing the transcriptions with an AR technique. Our experiment showed that the word error rate was improved by more than 10 points with the run-time adaptation using only twelve minutes observation.

Direct methods have shown excellent performance in the applications of visual odometry and SLAM. In this work we propose to leverage their effectiveness for the task of 3D multi-object tracking. To this end, we propose DirectTracker, a framework that effectively combines direct image alignment for the short-term tracking and sliding-window photometric bundle adjustment for 3D object detection. Object proposals are estimated based on the sparse sliding-window pointcloud and further refined using an optimization-based cost function that carefully combines 3D and 2D cues to ensure consistency in image and world space. We propose to evaluate 3D tracking using the recently introduced higher-order tracking accuracy (HOTA) metric and the generalized intersection over union sim-ilarity measure to mitigate the limitations of the conventional use of intersection over union for the evaluation of vision-based trackers. We perform evaluation on the KITTI Tracking benchmark for the Car class and show competitive performance in tracking objects both in 2D and 3D.

A central aspect of robotic motion planning is collision avoidance, where a multitude of different approaches are currently in use. Optimization-based motion planning is one method, that often heavily relies on distance computations between robots and obstacles. These computations can easily become a bottleneck, as they do not scale well with the complexity of the robots or the environment. To improve performance, many different methods suggested to use collision primitives, i.e. simple shapes that approximate the more complex rigid bodies, and that are simpler to compute distances to and from. However, each pair of primitives requires its own specialized code, and certain pairs are known to suffer from numerical issues. In this paper, we propose an easy-to-use, unified treatment of a wide variety of primitives. We formulate distance computation as a minimization problem, which we solve iteratively. We show how to take derivatives of this minimization problem, allowing it to be seamlessly integrated into a trajectory optimization method. We demonstrate that the resulting method can be used to plan smooth and collision-free paths based on a variety of single- and multi-robot scenarios with different obstacles.

Research in manipulation of deformable objects is typically conducted on a limited range of scenarios, because handling each scenario on hardware takes significant effort. Realistic simulators with support for various types of deformations and interactions have the potential to speed up experimentation with novel tasks and algorithms. However, for highly deformable objects it is challenging to align the output of a simulator with the behavior of real objects. Manual tuning is not intuitive, hence automated methods are needed. We view this alignment problem as a joint perception-inference challenge and demonstrate how to use recent neural network architectures to successfully perform simulation parameter inference from real point clouds. We analyze the performance of various architectures, comparing their data and training requirements. Furthermore, we propose to leverage differentiable point cloud sampling and differentiable simulation to significantly reduce the time to achieve the alignment. We employ an efficient way to propagate gradients from point clouds to simulated meshes and further through to the physical simulation parameters, such as mass and stiffness. Experiments with highly deformable objects show that our method can achieve comparable or better alignment with real object behavior, while reducing the time needed to achieve this by more than an order of magnitude. Videos and supplementary material are available at https://tinyurl.com/diffcloud.

The observation of the anatomy of agile animals and their locomotion capabilities emphasizes the importance of fast and lightweight legs and confirms the intrinsic compliance integrated into muscle-tendon units as a major ingredient for energy efficient and robust locomotion. This quality is especially relevant for distal leg segments which are subject to aggressive dynamics. Legged robots are accordingly designed to improve dynamic performance by lightweight mechanisms combined with series elastic actuation systems. However, so far no designs are available that feature all characteristics of a perfect distal legged locomotion actuator such as a lightweight and low-inertia structure, with high mechanical efficiency, no stick and sliding friction, and low mechanical complexity. With this goal in mind, we propose a novel robotic leg which integrates all above features. Specifically, we develop, implement, and characterize a bioinspired robot leg that features a lightweight Series ELastic Diaphragm distal Actuator (SELDA) for active control of foot motion. We conducted experiments to compare two leg configurations, with and without foot actuation, to demonstrate the effectiveness of the proposed solution in agile forward hopping controlled by a central pattern generator. We studied how tuning SELDA's activation timing can adjust the robot's hopping height by 11% and its forward velocity by 14%, even with comparatively low power injection to the distal joint.