ICRA 2023 论文列表

Detecting obstacles is crucial for safe and efficient autonomous driving. To this end, we present NVRadarNet, a deep neural network (DNN) that detects dynamic obstacles and drivable free space using automotive RADAR sensors. The network utilizes temporally accumulated data from multiple RADAR sensors to detect dynamic obstacles and compute their orientation in a top-down bird's-eye view (BEV). The network also regresses drivable free space to detect unclassified obstacles. Our DNN is the first of its kind to utilize sparse RADAR signals in order to perform obstacle and free space detection in real time from RADAR data only. The network has been successfully used for perception on our autonomous vehicles in real self-driving scenarios. The network runs faster than real time on an embedded GPU and shows good generalization across geographic regions.11Video at https://youtu.be/WlwJJMltoJY.

In this work, we study how to build socially intelligent robots to assist people in their homes. In particular, we focus on assistance with online goal inference, where robots must simultaneously infer humans' goals and how to help them achieve those goals. Prior assistance methods either lack the adaptivity to adjust helping strategies (i.e., when and how to help) in response to uncertainty about goals or the scalability to conduct fast inference in a large goal space. Our NOPA (Neurally-guided Online Probabilistic Assistance) method addresses both of these challenges. NOPA consists of (1) an online goal inference module combining neural goal proposals with inverse planning and particle filtering for robust inference under uncertainty, and (2) a helping planner that discovers valuable subgoals to help with and is aware of the uncertainty in goal inference. We compare NOPA against multiple baselines in a new embodied AI assistance challenge: Online Watch-And-Help, in which a helper agent needs to simultaneously watch a main agent's action, infer its goal, and help perform a common household task faster in realistic virtual home environments. Experiments show that our helper agent robustly updates its goal inference and adapts its helping plans to the changing level of uncertainty.11Code and a supplementary video are available at https://www.tshu.io/online_watch_and_help.

There are a multitude of emerging imaging technologies that could benefit robotics. However the need for bespoke models, calibration and low-level processing represents a key barrier to their adoption. In this work we present NOCaL, Neural Odometry and Calibration using Light fields, a semi-supervised learning architecture capable of interpreting previously unseen cameras without calibration. NOCaL learns to estimate camera parameters, relative pose, and scene appearance. It employs a scene-rendering hypernetwork pre-trained on a large number of existing cameras and scenes, and adapts to previously unseen cameras using a small supervised training set to enforce metric scale. We demonstrate NOCaL on rendered and captured imagery using conventional cameras, demonstrating calibration-free odometry and novel view synthesis. This work represents a key step toward automating the interpretation of general camera geometries and emerging imaging technologies. Code and datasets are available at https://roboticimaging.org/Projects/NOCaL/.

We introduce NIFT, Neural Interaction Field and Template, a descriptive and robust interaction representation of object manipulations to facilitate imitation learning. Given a few object manipulation demos, NIFT guides the generation of the interaction imitation for a new object instance by matching the Neural Interaction Template (NIT) extracted from the demos in the target Neural Interaction Field (NIF) defined for the new object. Specifically, the NIF is a neural field that encodes the relationship between each spatial point and a given object, where the relative position is defined by a spherical distance function rather than occupancies or signed distances, which are commonly adopted by conventional neural fields but less informative. For a given demo interaction, the corresponding NIT is defined by a set of spatial points sampled in the demo NIF with associated neural features. To better capture the interaction, the points are sampled on the Interaction Bisector Surface (IBS), which consists of points that are equidistant to the two interacting objects and has been used extensively for interaction representation. With both point selection and pointwise features defined for better interaction encoding, NIT effectively guides the feature matching in the NIFs of the new object instances such that the relative poses are optimized to realize the manipulation while imitating the demo interactions. Experiments show that our NIFT solution outperforms state-of-the-art imitation learning methods for object manipulation and generalizes better to objects from new categories.

Current research strive has tremendously changed the horizon of computer vision tasks in multiple agents tracking. Nevertheless, in the research of robotic assisted surgery, reliable surgical instrument tracking imposes challenge due to the high complexity in state modeling for the hierarchical structure of the instrument versus de-coupling the spatial-temporal correlations naturally embedded in the task. In this paper, we present a new tracking paradigm integrating the trajectory prediction to reduce the data association error that is propagated from the false detection. As a key component in the system, a proposed predictor disentangles the hierarchical modeling and agent kinematic learning by introducing inductive attention mechanism in spatial-temporal graph network. Experiments on real anatomical datasets show that our tracking-by-prediction scheme improves overall localization accuracy over the frames by up to 81%, in comparison to the generic pipelines of tracking, even with transductive graph representation learning, with a large margin of gain in terms of precise localization.

We have developed an autonomous robot motion generation model based on distributed and integrated multimodal learning. Since each modality used as a robot's senses, such as image, joint angle, and torque, has a different physical meaning and time characteristic, the generation of autonomous motions using multimodal learning has sometimes failed due to overlearning in one of the modalities. Inspired by the sensory processing of the human brain, our model is based on the processing of each sense performed in the primary somatosensory cortex and the integrated processing of multiple senses in the association cortex and the primary motor cortex. Specifically, the proposed model utilizes two types of recurrent neural networks: sensory RNNs, which learn each sense in a time series, and a union RNN, which communicates with sensory RNNs and learns sensory integration. The simulation results of multiple tasks showed that our model processes multiple modalities appropriately and generates smoother motions with lower jerk than the conventional model. We also demonstrated a chair assembly task by combining fixed motions and autonomous motions with our model.

This paper addresses the challenge of reconstructing a scene with a neural radiance field (NeRF) for robot vision and scene understanding using multiple modalities. Researchers have introduced the use of NeRF to represent an object for synthesizing and rendering novel views of complex scenes by optimizing a 3-D radiance field for ray casting and rendering for 2-D RGB images. However, using RGB images alone introduces additional geometry ambiguities with transparent objects or complex scenes and cannot accurately depict the 3-D shapes. We discuss and solve this problem and use multiple modalities as input for the same NeRF model to build a multimodal NeRF by incorporating point clouds and infrared image supervision to prevent such bias. In contrast to RGB images, infrared images and point clouds are typically taken by separate cameras that cannot be aligned with the RGB camera. We further introduce the alignment of different modalities based on point cloud registration to estimate the relative transformation matrices between them before training a NeRF model with multiple modalities. We evaluate our model on chosen scenes from the ScanNet and M2DGR datasets and demonstrate that it outperforms existing state-of-the-art methods.

Visual navigation plays an important role for Unmanned Aerial Vehicles(UAVs). In some applications, the landmark image and the real-time image may be heterogeneous, like near-infrared and visible images. In this work, we propose a multimodal image registration method to deal with near-infrared and visible images so that it can be applied to visual navigation system for the localization of UAVs in GPS-denied environments. At first, a new feature extraction strategy is developed to embed different modalities of images into the common feature space based on disentangled representations. Such common space is independent of the image modality, and this can eliminate the modality differences. Meanwhile, an intensity loss is introduced to measure the similarity of mono-modal images. In the proposed method, we can directly predict the transformation parameters and thus accelerates the localization of UAV s. Extensive experiments on synthetic datasets are conducted to demonstrate the validity of our method, and the experimental results show that the proposed method can effectively improve the localization accuracy.

We derive a learning framework to generate routing/pickup policies for a fleet of autonomous vehicles tasked with servicing stochastically appearing requests on a city map. We focus on policies that 1) give rise to coordination amongst the vehicles, thereby reducing wait times for servicing requests, 2) are non-myopic, and consider a-priori potential future requests, 3) can adapt to changes in the underlying demand distribution. Specifically, we are interested in policies that are adaptive to fluctuations of actual demand conditions in urban environments, such as on-peak vs. off-peak hours. We achieve this through a combination of (i) an online play algorithm that improves the performance of an offline-trained policy, and (ii) an offline approximation scheme that allows for adapting to changes in the underlying demand model. In particular, we achieve adaptivity of our learned policy to different demand distributions by quantifying a region of validity using the q-valid radius of a Wasserstein Ambiguity Set. We propose a mechanism for switching the originally trained offline approximation when the current demand is outside the original validity region. In this case, we propose to use an offline architecture, trained on a historical demand model that is closer to the current demand in terms of Wasserstein distance. We learn routing and pickup policies over real taxicab requests in San Francisco with high variability between on-peak and off-peak hours, demonstrating the ability of our method to adapt to real fluctuation in demand distributions. Our numerical results demonstrate that our method outperforms alternative rollout-based reinforcement learning schemes, as well as other classical methods from operations research.

In many automation tasks involving manipulation of rigid objects, the poses of the objects must be acquired. Vision-based pose estimation using a single RGB or RGB-D sensor is especially popular due to its broad applicability. However, single-view pose estimation is inherently limited by depth ambiguity and ambiguities imposed by various phenom-ena like occlusion, self-occlusion, reflections, etc. Aggregation of information from multiple views can potentially resolve these ambiguities, but the current state-of-the-art multi-view pose estimation method only uses multiple views to aggregate single-view pose estimates, and thus rely on obtaining good single-view estimates. We present a multi-view pose estimation method which aggregates learned 2D-3D distributions from multiple views for both the initial estimate and optional refinement. Our method performs probabilistic sampling of 3D-3D correspondences under epipolar constraints using learned 2D-3D correspondence distributions which are implicitly trained to respect visual ambiguities such as symmetry. Evaluation on the T-LESS dataset shows that our method reduces pose estimation errors by 80–91% compared to the best single-view method, and we present state-of-the-art results on T-LESS with four views, even compared with methods using five and eight views.

3D point cloud semantic segmentation is one of the fundamental tasks for environmental understanding. Although significant progress has been made in recent years, the performance of classes with few examples or few points is still far from satisfactory. In this paper, we propose a novel multi-to-single knowledge distillation framework for the 3D point cloud semantic segmentation task to boost the performance of those hard classes. Instead of fusing all the points of multi-scans directly, only the instances that belong to the previously defined hard classes are fused. To effectively and sufficiently distill valuable knowledge from multi-scans, we leverage a multilevel distillation framework, i.e., feature representation distillation, logit distillation, and affinity distillation. We further develop a novel instance-aware affinity distillation algorithm for capturing high-level structural knowledge to enhance the distillation efficacy for hard classes. Finally, we conduct experiments on the SemanticKITTI dataset, and the results on both the validation and test sets demonstrate that our method yields substantial improvements compared with the baseline method. The code is available at https://github.com/skyshoumeng/M2SKD.

High-dimensional expensive problems are often encountered in the design and optimization of complex robotic and automated systems and distributed computing systems, and they suffer from a time-consuming fitness evaluation process. It is extremely challenging and difficult to produce promising solutions in a high-dimensional search space. This work proposes an evolutionary optimization framework with embedded autoencoders that effectively solve optimization problems with high-dimensional search space. Autoencoders provide strong dimension reduction and feature extraction abilities that compress a high-dimensional space to an informative low-dimensional one. Search operations are performed in a low-dimensional space, thereby guiding whole population to converge to the optimal solution more efficiently. Multiple subpopulations coevolve iteratively in a distributed manner. One subpopulation is embedded by an autoencoder, and the other one is guided by a newly proposed Multi-swarm Gray-wolf-optimizer based on Genetic-learning (MGG). Thus, the proposed multi-swarm framework is named Autoencoder-based MGG (AMGG). AMGG consists of three proposed strategies that balance exploration and exploitation abilities, i.e., a dynamic subgroup number strategy for reducing the number of subpopulations, a subpopulation reorganization strategy for sharing useful information about each subpopulation, and a purposeful detection strategy for escaping from local optima and improving exploration ability. AMGG is compared with several widely used algorithms by solving benchmark problems and a real-life optimization one. The results well verify that AMGG outperforms its peers in terms of search accuracy and convergence efficiency.

The performance of road defect segmentation (a.k.a. pixel-level road defect detection) has been improved alongside with remarkable achievement of deep learning. Those improvements need a large-scale and well-constructed dataset. However, road surface materials or designs vary from country to country, and the patterns of defects are hard to pre-define. In this paper, we propose a novel multi-source domain adaptation method to boost the performance of road defect segmentation on an unlabelled dataset. The proposed method generates multi-source ensembled labels using transferred information from models trained with multiple labelled source domains, which are utilised as supervisory signals for the unlabelled target domain. Furthermore, to reduce the domain gap between each source domain and a target domain, these domains are re-aligned with outlier repositioning to improve the defect segmentation performance. We demonstrate the effectiveness of our proposed method on Cracktree200, CRACK500, CFD, and Crack360 datasets. Experimental results show that the proposed method outperforms the existing unsupervised road defect segmentation methods and achieves competitive performance compared with recent supervised methods. The source code is publicly available on https://github.com/andreYoo/MSDA_RDS.git.

Most legged robots are built with leg structures from serially mounted links and actuators and are controlled through complex controllers and sensor feedback. In comparison, animals developed multi-segment legs, mechanical coupling between joints, and multi-segmented feet. They run agile over all terrains, arguably with simpler locomotion control. Here we focus on developing foot mechanisms that resist slipping and sinking also in natural terrain. We present first results of multi-segment feet mounted to a bird-inspired robot leg with multi-joint mechanical tendon coupling. Our one- and two-segment, mechanically adaptive feet show increased viable horizontal forces on multiple soft and hard substrates before starting to slip. We also observe that segmented feet reduce sinking on soft substrates compared to ball-feet and cylinder-feet. We report how multi-segmented feet provide a large range of viable centre of pressure points well suited for bipedal robots, but also for quadruped robots on slopes and natural terrain. Our results also offer a functional understanding of segmented feet in animals like ratite birds.

Tactile sensing has been increasingly utilized in robot control of unknown objects to infer physical properties and optimize manipulation. However, there is limited understanding about the contribution of different sensory modalities during interactive perception in complex interaction both in robots and in humans. This study investigated the effect of visual and haptic information on humans' exploratory interactions with a ‘cup of coffee’, an object with nonlinear internal dynamics. Subjects were instructed to rhythmically transport a virtual cup with a rolling ball inside between two targets at a specified frequency, using a robotic interface. The cup and targets were displayed on a screen, and force feedback from the cup-and-ball dynamics was provided via the robotic manipulandum. Subjects were encouraged to explore and prepare the dynamics by “shaking” the cup-and-ball system to find the best initial conditions prior to the task. Two groups of subjects received the full haptic feedback about the cup-and-ball movement during the task; however, for one group the ball movement was visually occluded. Visual information about the ball movement had two distinctive effects on the performance: it reduced preparation time needed to understand the dynamics and, importantly, it led to simpler, more linear input-output interactions between hand and object. The results highlight how visual and haptic information regarding nonlinear internal dynamics have distinct roles for the interactive perception of complex objects.

Forecasting the future states of surrounding traffic participants is a crucial capability for autonomous vehicles. The recently proposed occupancy flow field prediction introduces a scalable and effective representation to jointly predict surrounding agents' future motions in a scene. However, the challenging part is to model the underlying social interactions among traffic agents and the relations between occupancy and flow. Therefore, this paper proposes a novel Multi-modal Hierarchical Transformer network that fuses the vectorized (agent motion) and visual (scene flow, map, and occupancy) modalities and jointly predicts the flow and occupancy of the scene. Specifically, visual and vector features from sensory data are encoded through a multi-stage Transformer module and then a late-fusion Transformer module with temporal pixel-wise attention. Importantly, a flow-guided multi-head self-attention (FG-MSA) module is designed to better aggregate the information on occupancy and flow and model the mathematical relations between them. The proposed method is comprehensively validated on the Waymo Open Motion Dataset and compared against several state-of-the-art models. The results reveal that our model with much more compact architecture and data inputs than other methods can achieve comparable performance. We also demonstrate the effectiveness of incorporating vectorized agent motion features and the proposed FG-MSA module. Compared to the ablated model without the FG-MSA module, which won 2 nd place in the 2022 Waymo Occupancy and Flow Prediction Challenge, the current model shows better separability for flow and occupancy and further performance improvements.

Robots are traditionally bounded by a fixed embodiment during their operational lifetime, which limits their ability to adapt to their surroundings. Co-optimizing control and morphology of a robot, however, is often inefficient due to the complex interplay between the controller and morphology. In this paper, we propose a learning-based control method that can inherently take morphology into consideration such that once the control policy is trained in the simulator, it can be easily deployed to real robots with different embodiments. In particular, we present the Embodiment-aware Transformer (EAT), an architecture that casts this control problem as conditional sequence modeling. EAT outputs the optimal actions by leveraging a causally masked Transformer. By conditioning an autoregressive model on the desired robot embodiment, past states, and actions, our EAT model can generate future actions that best fit the current robot embodiment. Experimental results show that EAT can outperform all other alternatives in embodiment-varying tasks, and succeed in an example of real-world evolution tasks: stepping down a stair through updating the morphology alone. We hope that EAT will inspire a new push toward real-world evolution across many domains, where algorithms like EAT can blaze a trail by bridging the field of evolutionary robotics and big data sequence modeling.

6D Object pose estimation is a fundamental component in robotics enabling efficient interaction with the environment. 6D pose estimation is particularly challenging in bin- picking applications, where many objects are low-feature and reflective, and self-occlusion between objects of the same type is common. We propose a novel multi-view approach leveraging known camera transformations from an eye-in-hand setup to combine heatmap and keypoint estimates into a probability density map over 3D space. The result is a robust approach that is scalable in the number of views. It relies on a confidence score composed of keypoint probabilities and point-cloud alignment error, which allows reliable rejection of false positives. We demonstrate an average pose estimation error of approximately 0.5 mm and 2 degrees across a variety of difficult low-feature and reflective objects in the ROBI dataset, while also surpassing the state-of-art correct detection rate, measured using the 10% object diameter threshold on ADD error.

Multi-agent pursuit-evasion tasks involving intelligent targets are notoriously challenging coordination problems. In this paper, we investigate new ways to learn such coordinated behaviors of unmanned aerial vehicles (UAVs) aimed at keeping track of multiple evasive targets. Within a Multi-Agent Reinforcement Learning (MARL) framework, we specifically propose a variant of the Multi-Agent Deep Deterministic Policy Gradient (MADDPG) method. Our approach addresses multi-target pursuit-evasion scenarios within non-stationary and unknown environments with random obstacles. In addition, given the critical role played by collective exploration in terms of detecting possible targets, we implement heterogeneous roles for the pursuers for enhanced exploratory actions balanced by exploitation (i.e. tracking) of previously identified targets. Our proposed role-based MADDPG algorithm is not only able to track multiple targets, but also is able to explore for possible targets by means of the proposed Voronoi-based rewarding policy. We implemented, tested and validated our approach in a simulation environment prior to deploying a real-world multi-robot system comprising of Crazyflie drones. Our results demonstrate that a multi-agent pursuit team has the ability to learn highly efficient coordinated control policies in terms of target tracking and exploration even when confronted with multiple fast evasive targets in complex environments.