IROS 2022 Papers

The feature pyramid, which is a vital component of the convolutional neural networks, plays a significant role in several perception tasks, including object detection for autonomous driving. However, how to better fuse multi-level and multi-sensor feature pyramids is still a significant challenge, especially for object detection. This paper presents a FocusTR (Focusing on the valuable features by multiple Transformers), which is a simple yet effective architecture, to fuse feature pyramid for the single-stream 2D detector and two-stream 3D detector. Specifically, FocusTR encompasses several novel self-attention mechanisms, including the spatial-wise boxAlign attention (SB) for low-level spatial locations, context-wise affinity attention (CA) for high-level context information, and level-wise attention for the multi-level feature. To alleviate self-attention's computational complexity and slow training convergence, Fo-cusTR introduces a low and high-level fusion (LHF) to reduce the computational parameters, and the Pre- Ln [1]to accelerate the training convergence.

Autonomous navigation among people is a com-plex problem that also exhibits considerable variation depending on the type of environment and people involved. Here we consider navigation among crowds that exhibit flow-like behavior like people moving through a train station. We propose a novel pseudo-fluid model of crowd flow for such problems. These have an intuitive physical interpretation and do not require much tuning. We further formalize an observation model to infer flow properties from discrete sensor observations, including support for partial observability, and pair it with a flow-aware planner. We demonstrate the potential of the approach in simulated navigation scenarios. We achieve state of the art results on the CrowdBot navigation benchmark, and also compare favorably against a standard ROS planner on a partially observable environment, demonstrating that the flow-aware planner successfully estimates and plans around counter-flows in the crowd in real time. We conclude that flow-based planning shows great promise for crowded environments that may exhibit such flow-like behavior.

Processing large indoor scenes is a challenging task, as scan registration and camera trajectory estimation methods accumulate errors across time. As a result, the quality of reconstructed scans is insufficient for some applications, such as visual-based localization and navigation, where the correct position of walls is crucial. For many indoor scenes, there exists an image of a technical ftoorplan that contains information about the geometry and main structural elements of the scene, such as walls, partitions, and doors. We argue that such a ftoorplan is a useful source of spatial information, which can guide a 3D model optimization. The standard RGB-D 3D reconstruction pipeline consists of a tracking module applied to an RGB-D sequence and a bundle adjustment (BA) module that takes the posed RGB-D sequence and corrects the camera poses to improve consistency. We propose a novel optimization algorithm expanding conventional BA that leverages the prior knowledge about the scene structure in the form of a ftoorplan. Our experiments on the Redwood dataset and our self-captured data demonstrate that utilizing ftoorplan improves accuracy of 3D reconstructions.

Floor plans are the basis of reasoning in and communicating about indoor environments. In this paper, we show that by modelling floor plans as sequences of line segments seen from a particular point of view, recent advances in autoregressive sequence modelling can be leveraged to model and predict floor plans. The line segments are canonicalized and translated to sequence of tokens and an attention-based neural network is used to fit a one-step distribution over next tokens. We fit the network to sequences derived from a set of large-scale floor plans, and demonstrate the capabilities of the model in four scenarios: novel floor plan generation, completion of partially observed floor plans, generation of floor plans from simulated sensor data, and finally, the applicability of a floor plan model in predicting the shortest distance with partial knowledge of the environment.

Snap-fit peg-in-hole assembly widely exists in both industry and daily life, especially for consumer electronics. The buckle mechanism leads to a damping zone inside the port where insertion force needs to be increased. It is much difficult to automate this process by robots, for size and clearance of the components are always small, and the damping buckle should be perceived and distinguished from solid inner walls of the port. End-effector position control might be invalid, since grasping error will make it difficult to locate the plug accurately. In this article, we undertake this assembly challenge by taking advantage of fingertip tactile perception combined with visual images and force feedback. Raw sensor data is collected, processed, and fused together to be state input of a reinforcement learning network, generating continuous action vectors. We also propose a novel damping zone predictor through feature extraction and multimodal fusion, which is able to identify whether the plug has touched the buckle mechanism, so as to adjust the insertion force. The whole framework is implemented through a common USB Type-C insertion experiment on Franka Panda robot platform, reaching a success rate of 88%. Furthermore, system robustness is verified, and comparisons of different modalities are also conducted.

This paper addresses the flexible formation problem for unmanned surface vehicles in the presence of obstacles. Building upon the leader-follower formation scheme, a hybrid line-of-sight based flexible platooning method is proposed for follower vehicle to keep tracking the leader ship. A fusion artificial potential field collision avoidance approach is tailored to generate optimal collision-free trajectories for the vehicle to track. To steer the vehicle towards and stay within the neighborhood of the generated collision-free trajectory, a nonlinear model predictive controller is designed. Experimental results are presented to validate the efficiency of proposed method, showing that the unmanned surface vehicle is able to track the leader ship without colliding with the surrounded static obstacles in the considered experiments.

Motion prediction of road users in traffic scenes is critical for autonomous driving systems that must take safe and robust decisions in complex dynamic environments. We present a novel motion prediction system for autonomous driving. Our system is based on the Bayesian inverse planning framework, which efficiently orchestrates map-based goal extraction, a classical control-based trajectory generator and a mixture of experts collection of light-weight neural networks specialised in motion profile prediction. In contrast to many alternative methods, this modularity helps isolate performance factors and better interpret results, without compromising performance. This system addresses multiple aspects of interest, namely multi-modality, motion profile uncertainty and trajectory physical feasibility. We report on several experiments with the popular highway dataset NGSIM, demonstrating state-of-the-art performance in terms of trajectory error. We also perform a detailed analysis of our system's components, along with experiments that stratify the data based on behaviours, such as change-lane versus follow-lane, to provide insights into the challenges in this domain. Finally, we present a qualitative analysis to show other benefits of our approach, such as the ability to interpret the outputs.

This paper expands on the problem of grasping an object that can only be grasped by a single parallel gripper when a fixture (e.g., wall, heavy object) is harnessed. Preceding work that tackle this problem are limited in that the employed networks implicitly learn specific targets and fixtures to leverage. However, the notion of a usable fixture can vary in different environments, at times without any outwardly noticeable differences. In this paper, we propose a method to relax this limitation and further handle environments where the fixture location is unknown. The problem is formulated as visual affordance learning in a partially observable setting. We present a self-supervised reinforcement learning algorithm, Fixture-Aware Double Deep Q-Network (FA-DDQN), that processes the scene observation to 1) identify the target object based on a reference image, 2) distinguish possible fixtures based on interaction with the environment, and finally 3) fuse the information to generate a visual affordance map to guide the robot to successful Slide-to-Wall grasps. We demonstrate our proposed solution in simulation and in real robot experiments to show that in addition to achieving higher success than baselines, it also performs zero-shot generalization to novel scenes with unseen object configurations.

Fisheye object detection is a difficult task in robotics and autonomous driving. One of the reasons is that the fisheye datasets are inferior to standard image datasets in scale and quantity, which inspires the idea of using standard image datasets for fisheye object detection. However, the models trained on standard image datasets do not perform well with fisheye data. In this work, we explore the effect of fisheye images on different stages of the YOLOX with published weights generated by standard image datasets. We also propose a new regression strategy for 24-points object representation method, which is insensitive to image distortion. The experiments show that the feature extraction part is robust to fisheye image features, while the regression part of location and category performs poorly. The strategy can achieve the position of discrete points without calculating the IOU of irregular-shaped boxes. Theoretically, the strategy can be widely adopted to regress the irregular bounding boxes composed of discrete points. Source code is at https://github.com/IN2-ViAUn/Exploration-of-Potential.

Measuring the fidelity of synthesized virtual traffic has become an important and fundamental concern for evaluating the performance of different traffic simulation techniques and applications of autonomous vehicle testing. In this work, we propose a novel method to evaluate the fidelity of any trajectory data from the perspective of anomalous trajectory detection. First, given the trajectory data to be evaluated as input, the method learns spatio-temporal traffic features and reconstructs the input trajectory through a Long Short-Term Memory (LSTM)-based autoencoder architecture. Then, the anomalous trajectories are detected by comparing the reconstructed trajectories and the input ones using the reconstruction error as the benchmark. Our method can detect eight different kinds of anomalous trajectory in terms of changes in velocity and moving direction. In order to evaluate the fidelity of the input trajectory, we design a perceptual evaluation on virtual traffic fidelity and derive a mapping from the reconstruction error to the evaluation score. We demonstrated the effectiveness and robustness of our metric through many experiments on real-world and synthetic trajectory data containing different types of motion anomalies.

Tactile feedback is necessary for closing the sen-sorimotor loop in prosthetic and tele-operable control, which would allow for more precise manipulation and increased acceptance of use of such devices. Pressure stimuli are commonly presented to users in haptic devices through a sensory substitution to vibration. The precise nature of this substitution affects pressure sensitivity, as well as the comfort and intuitiveness of the device for the user. This study determines the effects of different vibrational encodings for pressure on user-preference and performance in a 4-alternative absolute identification task. 4 different encoding patterns for pressure were examined: short pulse and long pulse amplitude modulation along with sine and square wave frequency modulation. Of the methods examined, frequency modulation methods led to the best discrimination of stimuli (p < 0.001). There was a notable difference in user preference between the two frequency modulated systems, with sinusoidal stimulation being highest ranked across all the preference metrics and square-wave being ranked lowest in two of the three. This difference trended towards, but did not achieve statistical significance (TLX rankings, ppp = 0.098). This suggests that prostheses or teleoperated devices utilising vibrotactile feedback may benefit from implementing a discrete frequency-based sinusoidal pattern to indicate changes in grip force.

Humans are able to manipulate Deformable Linear Objects (DLOs) such as cables and wires, with little or no visual information, relying mostly on force sensing. In this work, we propose a reduced DLO model which enables such blind manipulation by keeping the object under tension. Further, an online model estimation procedure is also proposed. A set of elementary sliding and clipping manipulation primitives are defined based on our model. The combination of these primitives allows for more complex motions such as winding of a DLO. The model estimation and manipulation primitives are tested individually but also together in a real-world cable harness production task, using a dual-arm YuMi, thus demonstrating that force-based perception can be sufficient even for such a complex scenario.

Socially aware robot navigation, where a robot is required to optimize its trajectory to maintain comfortable and compliant spatial interactions with humans in addition to reaching its goal without collisions, is a fundamental yet challenging task in the context of human-robot interaction. While existing learning-based methods have achieved better performance than the preceding model-based ones, they still have drawbacks: reinforcement learning depends on the handcrafted reward that is unlikely to effectively quantify broad social compliance, and can lead to reward exploitation problems; meanwhile, inverse rein-forcement learning suffers from the need for expensive human demonstrations. In this paper, we propose a feedback-efficient active preference learning approach, FAPL, that distills human comfort and expectation into a reward model to guide the robot agent to explore latent aspects of social compliance. We further introduce hybrid experience learning to improve the efficiency of human feedback and samples, and evaluate benefits of robot behaviors learned from FAPL through extensive simulation experiments and a user study (N=10) employing a physical robot to navigate with human subjects in real-world scenarios. Source code and experiment videos for this work are available at: https://sites.google.com/view/san-fapl.

Learning from Demonstration (LfD) is a very appealing approach to empower robots with autonomy. Given some demonstrations provided by a human teacher, the robot can learn a policy to solve the task without explicit programming. A promising use case is to endow smart robotic wheelchairs with active assistance to navigation. By using LfD, it is possible to learn to infer short-term destinations anywhere, without the need of building a map of the environment beforehand. Nevertheless, it is difficult to generalize robot behaviors to environments other than those used for training. We believe that one possible solution is learning from crowds, involving a broad number of teachers (the end users themselves) who perform demonstrations in diverse and real environments. To this end, in this work we consider Federated Learning from Demonstration (FLfD), a distributed approach based on a Federated Learning architecture. Our proposal allows the training of a global deep neural network using sensitive local data (images and laser readings) with privacy guarantees. In our experiments we pose a scenario involving different clients working in heterogeneous domains. We show that the federated model is able to generalize and deal with non Independent and Identically Distributed (non-IID) data.

Semantic Segmentation is essential to make self-driving vehicles autonomous, enabling them to understand their surroundings by assigning individual pixels to known categories. However, it operates on sensible data collected from the users' cars; thus, protecting the clients' privacy becomes a primary concern. For similar reasons, Federated Learning has been recently introduced as a new machine learning paradigm aiming to learn a global model while preserving privacy and leveraging data on millions of remote devices. Despite several efforts on this topic, no work has explicitly addressed the challenges of federated learning in semantic segmentation for driving so far. To fill this gap, we propose FedDrive, a new benchmark consisting of three settings and two datasets, incorporating the real-world challenges of statistical heterogeneity and domain generalization. We benchmark state-of-the-art algorithms from the federated learning literature through an in-depth analysis, combining them with style transfer methods to improve their generalization ability. We demonstrate that correctly handling normalization statistics is crucial to deal with the aforementioned challenges. Furthermore, style transfer improves performance when dealing with significant appearance shifts. Official website: https://feddrive.github.io.

Autonomously driving vehicles must be able to navigate in dynamic and unpredictable environments in a collision-free manner. So far, this has only been partially achieved in driverless cars and warehouse installations where marked structures such as roads, lanes, and traffic signs simplify the motion planning and collision avoidance problem. We are presenting a new control approach for car-like vehicles that is based on an unprecedentedly fast-paced A* implementation that allows the control cycle to run at a frequency of 30 Hz. This frequency enables us to place our A* algorithm as a low-level replanning controller that is well suited for navigation and collision avoidance in virtually any dynamic environment. Due to an efficient heuristic consisting of rotate-translate-rotate motions laid out along the shortest path to the target, our Short-Term Aborting A* (STAA*) converges fast and can be aborted early in order to guarantee a high and steady control rate. While our STAA* expands states along the shortest path, it takes care of collision checking with the environment including predicted states of moving obstacles, and returns the best solution found when the computation time runs out. Despite the bounded computation time, our STAA* does not get trapped in corners due to the following of the shortest path. In simulated and real-robot experiments, we demonstrate that our control approach eliminates collisions almost entirely and is superior to an improved version of the Dynamic Window Approach with predictive collision avoidance capabilities [1].

Autonomous exploration in unknown environments is a fundamental task for robots. Existing approaches mostly were concentrated on the efficiency of the exploration with the assumption of perfect state estimation, but the drift of pose estimation in visual SLAM occurs frequently and is detrimental to robot's localization and exploration performance. In this paper, a perception-aware exploration(PAE) method is proposed for rapidly and safely autonomous exploration in outdoor environments. The adaptive semantic information is proposed to improve the robustness of perception. Based on the perception module, both the selection of exploration goal on a novel weighted information gain and path planning can avoid the areas with high localization uncertainty. In addition, thanks to the proposed pipeline, including scan-based frontier detection, kd-tree based map prediction and suboptimal frontier buffer strategy, the PAE planner can explore the environment with high success rate and high efficiency. Several simulations are performed to verify the effectiveness of our methods.

Transferring tools and objects to human hands is an important ability of collaborative robots. Most of the existing approaches focus on handover affordance, however, the comfort of receiving objects with human hands is often neglected. In this paper, we use advanced deep learning models to pre-generate handover target configurations that are convenient for human grasping based on the characteristics of the objects and tools, and then the robot grasps and passes the objects to the human. Experimental results on a mobile collaborative robot show that our proposed framework can robustly and efficiently deliver different shapes and types of objects to a human hand of any pose within the robot's field of view in a target pose that is convenient for grasping and can quickly deliver objects to a new target location even after the human hand moves to a new position.

Perceptual Aliasing is one of the main problems in simultaneous localization and mapping (SLAM). Wrong associations between different places may lead to failure of the whole map. Research on structure information is rarely investigated among existing solutions to this problem. In cases of visual SLAM without sensors, such as LiDAR or Inertial Measurement Unit (IMU), structure information can rarely be obtained due to the sparsity of 3D points, which also makes structure analysis complex. This study provides a spherical harmonics (SH) based fast structural representation (SH-FS) in visual SLAM using sparse point clouds, which extracts the structure information from sparse points into single vector. SH-FS was applied in conventional feature-based loop closing process. Furthermore, a structure-aware loop closing method in visual SLAM was proposed to improve the robustness of SLAM systems. Moreover, our methods show a favorable performance in extensive experiments on different large-scale real world datasets.