ICRA 2024 论文列表

In Total Knee Replacement Arthroplasty (TKA), surgical robotics can provide image-guided navigation to fit implants with high precision. Its tracking approach highly relies on inserting bone pins into the bones tracked by the optical tracking system. This is normally done by invasive, radiative manners (implantable markers and CT scans), which introduce unnecessary trauma and prolong the preparation time for patients. To tackle this issue, ultrasound-based bone tracking could offer an alternative. In this study, we proposed a novel deep-learning structure to improve the accuracy of bone tracking by an A-mode ultrasound (US). We first obtained a set of ultrasound dataset from the cadaver experiment, where the ground truth locations of bones were calculated using bone pins. These data were used to train the proposed CasAtt-UNet to predict bone location automatically and robustly. The ground truth bone locations and those locations of US were recorded simultaneously. Therefore, we could label bone peaks in the raw US signals. As a result, our method achieved sub-millimeter precision across all eight bone areas with the only exception of one channel in the ankle. This method enables the robust measurement of lower extremity bone positions from 1D raw ultrasound signals. It shows great potential to apply A-mode ultrasound in orthopedic surgery from safe, convenient, and efficient perspectives.

In order to navigate safely and reliably in novel environments, robots must estimate perceptual uncertainty when confronted with out-of-distribution (OOD) obstacles not seen in training data. We present a method to accurately estimate pixel-wise uncertainty in semantic segmentation without requiring real or synthetic OOD examples at training time. From a shared per-pixel latent feature representation, a classification network predicts a categorical distribution over semantic labels, while a normalizing flow estimates the probability density of features under the training distribution. The label distribution and density estimates are combined in a Dirichlet-based evidential uncertainty framework that efficiently computes epistemic and aleatoric uncertainty in a single neural network forward pass. Our method is enabled by three key contributions. First, we simplify the problem of learning a transformation to the training data density by starting from a fitted Gaussian mixture model instead of the conventional standard normal distribution. Second, we learn a richer and more expressive latent pixel representation to aid OOD detection by training a decoder to reconstruct input image patches. Third, we perform theoretical analysis of the loss function used in the evidential uncertainty framework and propose a principled objective that more accurately balances training the classification and density estimation networks. We demonstrate the accuracy of our uncertainty estimation approach under long-tail OOD obstacle classes for semantic segmentation in both off-road and urban driving environments.

Control policies trained using deep reinforcement learning often generate stiff, high-frequency motions in response to unexpected disturbances. To promote more natural and compliant balance recovery strategies, we propose a simple modification to the typical reinforcement learning training process. Our key insight is that stiff responses to perturbations are due to an agent’s incentive to maximize task rewards at all times, even as perturbations are being applied. As an alternative, we introduce an explicit recovery stage where tracking rewards are given irrespective of the motions generated by the control policy. This allows agents a chance to gradually recover from disturbances before attempting to carry out their main tasks. Through an in-depth analysis, we highlight both the compliant nature of the resulting control policies, as well as the benefits that compliance brings to legged locomotion. In our simulation and hardware experiments, the compliant policy achieves more robust, energy-efficient, and safe interactions with the environment.

What makes generalization hard for imitation learning in visual robotic manipulation? This question is difficult to approach at face value, but the environment from the perspective of a robot can often be decomposed into enumerable factors of variation, such as the lighting conditions or the placement of the camera. Empirically, generalization to some of these factors have presented a greater obstacle than others, but existing work sheds little light on precisely how much each factor contributes to the generalization gap. Towards an answer to this question, we study imitation learning policies in simulation and on a real robot language-conditioned manipulation task to quantify the difficulty of generalization to different (sets of) factors. We design a simulated benchmark of 19 tasks with 11 factors of variation to facilitate more controlled evaluations of generalization. From our study, we determine an ordering of factors based on generalization difficulty, that is consistent across simulation and our real robot setup.1

In a Human-in-the-Loop paradigm, a robotic agent is able to act mostly autonomously in solving a task, but can request help from an external expert when needed. However, knowing when to request such assistance is critical: too few requests can lead to the robot making mistakes, but too many requests can overload the expert. In this paper, we present a Reinforcement Learning based approach to this problem, where a semi-autonomous agent asks for external assistance when it has low confidence in the eventual success of the task. The confidence level is computed by estimating the variance of the return from the current state. We show that this estimate can be iteratively improved during training using a Bellman-like recursion. On discrete navigation problems with both fully-and partially-observable state information, we show that our method makes effective use of a limited budget of expert calls at run-time, despite having no access to the expert at training time.

Herding is performed by people or trained animals to control the movement of livestock under the desired direction of an operator. This paper presents a novel decentralized control strategy for a group of robots to herd animals which consists of two phases, a surrounding phase and a driving phase. In the surrounding phase, a custom artificial potential field is employed to simultaneously guide the robots to encircle the herd by tracking the outmost animals and maintaining a safe distance from other neighboring robots. Once the encirclement is complete, the robots transition to drive the animals toward a designated goal by simply maintaining their initial formation and traversing to it. Unlike existing works on herding using flocking control, local observations of the nearest animals and communication with other robots within the sensing range are the only requirements for the robots to surround and herd the animals effectively. Moreover, the animal-robot behavior model resembles the interaction of livestock in the presence of an external predatory threat, where robots act as predators. An analytical proof and empirical results collected from different simulators demonstrate that the proposed control enables the robots to converge around the boundary of the animals and guide them toward the designated goal.

Collision avoidance algorithms for Autonomous Surface Vehicles (ASV) that follow the Convention on the International Regulations for Preventing Collisions at Sea (COLREGs) have been proposed in recent years. However, it may be difficult and unsafe to follow COLREGs in congested waters, where multiple ASVs are navigating in the presence of static obstacles and strong currents, due to the complex interactions. To address this problem, we propose a decentralized multi-ASV collision avoidance policy based on Distributional Reinforcement Learning, which considers the interactions among ASVs as well as with static obstacles and current flows. We evaluate the performance of the proposed Distributional RL based policy against a traditional RL-based policy and two classical methods, Artificial Potential Fields (APF) and Reciprocal Velocity Obstacles (RVO), in simulation experiments, which show that the proposed policy achieves superior performance in navigation safety, while requiring minimal travel time and energy. A variant of our framework that automatically adapts its risk sensitivity is also demonstrated to improve ASV safety in highly congested environments.

This paper proposes a decentralized trajectory planning framework for the collision avoidance problem of multiple micro aerial vehicles (MAVs) in environments with static and dynamic obstacles. The framework utilizes spatiotemporal occupancy grid maps (SOGM), which forecast the occupancy status of neighboring space in the near future, as the environment representation. Based on this representation, we extend the kinodynamic A* and the corridor-constrained trajectory optimization algorithms to efficiently tackle static and dynamic obstacles with arbitrary shapes. Collision avoidance between communicating robots is integrated by sharing planned trajectories and projecting them onto the SOGM. The simulation results show that our method achieves competitive performance against state-of-the-art methods in dynamic environments with different numbers and shapes of obstacles. Finally, the proposed method is validated in real experiments.

Multi-agent multi-target tracking has a wide range of applications, including wildlife patrolling, security surveillance or environment monitoring. Such algorithms often make restrictive assumptions: the number of targets and/or their initial locations may be assumed known, or agents may be pre-assigned to monitor disjoint partitions of the environment, reducing the burden of exploration. This also limits applicability when there are fewer agents than targets, since agents are unable to continuously follow the targets in their fields of view. Multi-agent tracking algorithms additionally assume inter-agent synchronization of observations, or the presence of a central controller to coordinate joint actions. Instead, we focus on the setting of decentralized multi-agent, multi-target, simultaneous active search-and-tracking with asynchronous interagent communication. Our proposed algorithm DecSTER uses a sequential monte carlo implementation of the Probability Hypothesis Density filter for posterior inference combined with Thompson sampling for decentralized multi-agent decision making. We compare different action selection policies, focusing on scenarios where targets outnumber agents. In simulation, we demonstrate that DecSTER is robust to unreliable inter-agent communication and outperforms information-greedy baselines in terms of the Optimal Sub-Pattern Assignment (OSPA) metric for different numbers of targets and varying teamsizes.

Path planning for multiple non-holonomic robots in continuous domains constitutes a difficult robotics challenge with many applications. Despite significant recent progress on the topic, computationally efficient and high-quality solutions are lacking, especially in lifelong settings where robots must continuously take on new tasks. In this work, we make it possible to extend key ideas enabling state-of-the-art (SOTA) methods for multi-robot planning in discrete domains to the motion planning of multiple Ackerman (car-like) robots in lifelong settings, yielding high-performance centralized and decentralized planners. Our planners compute trajectories that allow the robots to reach precise SE(2) goal poses. The effectiveness of our methods is thoroughly evaluated and confirmed using both simulation and real-world experiments.

Scene flow estimation determines a scene’s 3D motion field, by predicting the motion of points in the scene, especially for aiding tasks in autonomous driving. Many networks with large-scale point clouds as input use voxelization to create a pseudo-image for real-time running. However, the voxelization process often results in the loss of point-specific features. This gives rise to a challenge in recovering those features for scene flow tasks. Our paper introduces DeFlow which enables a transition from voxel-based features to point features using Gated Recurrent Unit (GRU) refinement. To further enhance scene flow estimation performance, we formulate a novel loss function that accounts for the data imbalance between static and dynamic points. Evaluations on the Argoverse 2 scene flow task reveal that DeFlow achieves state-of-the-art results on large-scale point cloud data, demonstrating that our network has better performance and efficiency compared to others. The code is available at https://github.com/KTH-RPL/deflow.

This paper presents a novel framework for learning robust bipedal walking by combining a data-driven state representation with a Reinforcement Learning (RL) based locomotion policy. The framework utilizes an autoencoder to learn a low-dimensional latent space that captures the complex dynamics of bipedal locomotion from existing locomotion data. This reduced dimensional state representation is then used as states for training a robust RL-based gait policy, eliminating the need for heuristic state selections or the use of template models for gait planning. The results demonstrate that the learned latent variables are disentangled and directly correspond to different gaits or speeds, such as moving forward, backward, or walking in place. Compared to traditional template model-based approaches, our framework exhibits superior performance and robustness in simulation. The trained policy effectively tracks a wide range of walking speeds and demonstrates good generalization capabilities to unseen scenarios.

Recent deep learning based visual simultaneous localization and mapping (SLAM) methods have made significant progress. However, how to make full use of visual information as well as better integrate with inertial measurement unit (IMU) in visual SLAM has potential research value. This paper proposes a novel deep SLAM network with dual visual factors. The basic idea is to integrate both photometric factor and re-projection factor into the end-to-end differentiable structure through multi-factor data association module. We show that the proposed network dynamically learns and adjusts the confidence maps of both visual factors and it can be further extended to include the IMU factors as well. Extensive experiments validate that our proposed method significantly outperforms the state-of-the-art methods on several public datasets, including TartanAir, EuRoC and ETH3D-SLAM. Specifically, when dynamically fusing the three factors together, the absolute trajectory error for both monocular and stereo configurations on EuRoC dataset has reduced by 45.3% and 36.2% respectively.

Motion prediction and cost evaluation are vital components in the decision-making system of autonomous vehicles. However, existing methods often ignore the importance of cost learning and treat them as separate modules. In this study, we employ a tree-structured policy planner and propose a differentiable joint training framework for both ego-conditioned prediction and cost models, resulting in a direct improvement of the final planning performance. For conditional prediction, we introduce a query-centric Transformer model that performs efficient ego-conditioned motion prediction. For planning cost, we propose a learnable context-aware cost function with latent interaction features, facilitating differentiable joint learning. We validate our proposed approach using the real-world nuPlan dataset and its associated planning test platform. Our framework not only matches state-of-the-art planning methods but outperforms other learning-based methods in planning quality, while operating more efficiently in terms of runtime. We show that joint training delivers significantly better performance than separate training of the two modules. Additionally, we find that tree-structured policy planning outperforms the conventional single-stage planning approach. Code is available: https://github.com/MCZhi/DTPP.

An accurate motion model is a fundamental component of most autonomous navigation systems. While much work has been done on improving model formulation, no standard protocol exists for gathering empirical data required to train models. In this work, we address this issue by proposing Data-driven Robot Input Vector Exploration (DRIVE), a protocol that enables characterizing uncrewed ground vehicles (UGVs) input limits and gathering empirical model training data. We also propose a novel learned slip approach outperforming similar acceleration learning approaches. Our contributions are validated through an extensive experimental evaluation, cumulating over 7km and 1.8h of driving data over three distinct UGVs and four terrain types. We show that our protocol offers increased predictive performance over common human-driven data-gathering protocols. Furthermore, our protocol converges with 46 s of training data, almost four times less than the shortest human dataset gathering protocol. We show that the operational limit for our model is reached in extreme slip conditions encountered on surfaced ice. DRIVE is an efficient way of characterizing UGV motion in its operational conditions. Our code and dataset are both available online at this link: https://github.com/norlab-ulaval/DRIVE.

We propose a new system named DOS®(Deployment Operating System for Robots) for reliably deploying any data-driven robots in both production and simulation environments. Compared to existing systems, DOS® features a unique CI/CD (continuous integration and continuous deployment) architecture which allows us to seamlessly integrate agile development and reliable operation in a fully automated fashion. With this CI/CD architecture, this paper mainly introduces three essential components that uniquely differentiate DOS® from existing robotic systems: (i) An environment adapter that provides a systematic and robust approach to handle the deployment complexity in real world environments; (ii) A data replay reservoir that provides a unified data model supporting arbitrary robotic decision models; (iii) An analytical profiler that collects any set of user-defined performance metrics for system optimization. DOS® significantly increases the reliability and maintainability of the deployed robotic systems. To illustrate this point, we compare DOS® with more traditional approaches on deploying a navigational robot in a challenging working environment with many new corner case scenarios. Our results show that DOS® outperforms traditional approach in great magnitudes in terms of deployment time and operational robustness.

We propose a new method for fine registering multiple point clouds simultaneously. The approach is characterized by being dense, therefore point clouds are not reduced to pre-selected features in advance. Furthermore, the approach is robust against small overlaps and dynamic objects, since no direct correspondences are assumed between point clouds. Instead, all points are merged into a global point cloud, whose scattering is then iteratively reduced. This is achieved by dividing the global point cloud into uniform grid cells whose contents are subsequently modeled by normal distributions. We show that the proposed approach can be used in a sliding window continuous trajectory optimization combined with IMU measurements to obtain a highly accurate and robust LiDAR inertial odometry estimation. Furthermore, we show that the proposed approach is also suitable for large scale keyframe optimization to increase accuracy. We provide the source code and some experimental data on https://github.com/davidskdds/DMSA_LiDAR_SLAM.git.

Accurate pose estimation over SE(3) is fundamentally crucial for numerous perception tasks, including camera re-localization. While existing learning-based methods estimated from a series of RGB images have significantly improved the accuracy of pose, the majority of models still face one or two limitations. First, few representations on SE(3) are smooth and differential, making them difficult to apply in deep learning frameworks. Second, they often require high computational resources due to complex deep network designs. We in this paper propose the DL-PoseNet to address these issues. Specifically, we present a novel representation for SE(3) which follows the property of smoothness of the pose. We then design a lightweight neural network to regress the pose by developing a differential pose layer. Finally, we introduce a novel loss function and gradient descent method to better supervise the proposed lightweight pose network. Extensive experiments on the camera re-localization task on the Cambridge Landmarks and 7-Scenes datasets demonstrate the superior predictive accuracy and benefits of our method in comparison with the state-of-the-art.

This paper introduces a novel sonar odometry system that estimates the relative spatial transformation between two sonar image frames. Considering the unique challenges, such as low resolution and high noise, of sonar imagery for odometry and Simultaneous Localization and Mapping (SLAM), the proposed Direct Imaging Sonar Odometry (DISO) system is designed to estimate the relative transformation between two sonar frames by minimizing the aggregated sonar intensity errors of points with high intensity gradients. Moreover, DISO is implemented to incorporate a multi-sensor window optimization technique, a data association strategy and an acoustic intensity outlier rejection algorithm for reliability and accuracy. The effectiveness of DISO is evaluated using both simulated and real-world sonar datasets, showing that it outperforms the existing geometric-only method on localization accuracy and achieves state-of-the-art sonar odometry performance. We release the source codes of the DISO implementation to the community. The source code is available at https://github.com/SenseRoboticsLab/DISO.