IROS 2023 Papers

Occlusion remains an issue in multiple object tracking, which could cause ambiguity in object detection, such as incorrect or missing detection. Under occlusion, a track could experience an early termination, resulting in identity switches and/or fragmentation. To recover from different lengths of occlusions, the track should be maintained by considering its occlusion status. To address the issues mentioned above, we propose an indicator that can model the track's occlusion extent via geometric information provided by LiDAR data. Through incorporating the indicator into the track management and data association process, it is feasible to prevent tracks from premature termination. The proposed method is evaluated on the collected dataset which undergoes frequent and severe occlusions. Compared to the state-of-the-art probabilistic tracking approach, our approach achieves improvements of 3.26% in MOTA and 5.36% in IDF1. Additionally, we obtain 9.89% improvements in IDF1 specifically for objects experiencing severe occlusions.

State-of-the-art lidar panoptic segmentation (LPS) methods follow “bottom-up” segmentation-centric fashion wherein they build upon semantic segmentation networks by utilizing clustering to obtain object instances. In this paper, we re-think this approach and propose a surprisingly simple yet effective detection-centric network for both LPS and tracking. Our network is modular by design and optimized for all aspects of both the panoptic segmentation and tracking task. One of the core components of our network is the object instance detection branch, which we train using point-level (modal) annotations, as available in segmentation-centric datasets. In the absence of amodal (cuboid) annotations, we regress modal centroids and object extent using trajectory-level supervision that provides information about object size, which cannot be inferred from single scans due to occlusions and the sparse nature of the lidar data. We obtain fine-grained instance segments by learning to associate lidar points with detected centroids. We evaluate our method on several 3D/4D LPS benchmarks and observe that our model establishes a new state-of-the-art among open-sourced models, outperforming recent query-based models.

Most of the existing moving object segmentation (MOS) methods regard MOS as an independent task, in this paper, we associate the MOS task with semantic segmentation, and propose a semantics-guided network for moving object segmentation (LiDAR-SGMOS). We first transform the range image and semantic features of the past scan into the range view of current scan based on the relative pose between scans. The residual image is obtained by calculating the normalized absolute difference between the current and transformed range images. Then, we apply a Meta-Kernel based cross scan fusion (CSF) module to adaptively fuse the range images and semantic features of current scan, the residual image and transformed features. Finally, the fused features with rich motion and semantic information are processed to obtain reliable MOS results. We also introduce a residual image augmentation method to further improve the MOS performance. Our method outperforms most LiDAR-MOS methods with only two sequential LiDAR scans as inputs on the SemanticKITTI MOS dataset.

This paper proposes a LiDAR-Inertial SLAM with efficiently extracted planes, which couples explicit planes in the odometry to improve accuracy and in the mapping for consistency. The proposed method consists of three parts: an efficient Point \boldsymbol{\rightarrow\text{Line}\rightarrow \text{Plane}}\boldsymbol{\rightarrow\text{Line}\rightarrow \text{Plane}} extraction algorithm, a LiDAR-Inertial-Plane tightly coupled odometry, and a global plane-aided mapping. Specifically, we leverage the ring field of the LiDAR point cloud to accelerate the region-growing-based plane extraction algorithm. Then we tightly coupled IMU pre-integration factors, LiDAR odometry factors, and explicit plane factors in the sliding window to obtain a more accurate initial pose for mapping. Finally, we maintain explicit planes in the global map, and enhance system consistency by optimizing the factor graph of optimized odometry factors and plane observation factors. Experimental results show that our plane extraction method is efficient, and the proposed plane-aided LiDAR-Inertial SLAM significantly improves the accuracy and consistency compared to the other state-of-the-art algorithms with only a small increase in time consumption.

Autonomous driving in rain remains challenging. Rain causes sensor performance degradation that can affect sensor measurement quality. During the rain, lasers may suffer from energy loss due to raindrop absorption. As a result, some laser measurements reflected from obstacles may not be recognized by the LiDAR sensor, thus raising potential risks for autonomous vehicles. This work investigates a novel task that aims to detect those missing measurements. Our solution uses a two-stage learning method to generate an anomaly score for each missing measurement, representing the likelihood of being caused by rain. We evaluate our method with real-world data and demonstrate its effectiveness in identifying anomalous missing measurements through qualitative and quantitative experiments.

Depth estimation is one of the essential tasks to be addressed when creating mobile autonomous systems. While monocular depth estimation methods have improved in recent times, depth completion provides more accurate and reliable depth maps by additionally using sparse depth information from other sensors such as LiDAR. However, current methods are specifically trained for a single LiDAR sensor. As the scanning pattern differs between sensors, every new sensor would require re- training a specialized depth completion model, which is computationally inefficient and not flexible. Therefore, we propose to dynamically adapt the depth completion model to the used sensor type enabling LiDAR adaptive depth completion. Specifically, we propose a meta depth completion network that uses data patterns derived from the data to learn a task network to alter weights of the main depth completion network to solve a given depth completion task effectively. The method demonstrates a strong capability to work on multiple LiDAR scanning patterns and can also generalize to scanning patterns that are unseen during training. While using a single model, our method yields significantly better results than a non-adaptive baseline trained on different LiDAR patterns. It outperforms LiDAR-specific expert models for very sparse cases. These advantages allow flexible deployment of a single depth completion model on different sensors, which could also prove valuable to process the input of nascent LiDAR technology with adaptive instead of fixed scanning patterns. The source code is available at github.com/wbkit/ResLAN

Multi-goal motion planning requires a robot to plan collision-free and dynamically-feasible motions to reach multiple goals, often in unstructured, obstacle-rich environments. This is challenging due to the complex dependencies between navigation and high-level reasoning, requiring the robot to explore a vast space of feasible motions and goal sequences. Our approach combines machine learning and Traveling Salesman Problem (TSP) solvers with sampling-based motion planning. Machine learning predicts distances and directions between locations, considering obstacles and robot dynamics, which the TSP solver uses to compute promising tours. Sampling-based motion planning expands a motion tree to follow the tours along the predicted directions. We demonstrate the effectiveness of our approach through experiments with vehicle and snake-like robot models operating in unstructured environments with multiple goals.

As robots become increasingly prevalent amidst diverse environments, their ability to adapt to novel scenarios and objects is essential. Advances in modern object detection have also paved the way for robots to identify interaction entities within their immediate vicinity. One drawback is that the robot's operational domain must be known at the time of training, which hinders the robot's ability to adapt to unexpected environments outside the preselected classes. However, when encountering such challenges a human can provide support to a robot by teaching it about the new, yet unknown objects on an ad hoc basis. In this work, we merge augmented reality and human gaze in the context of multimodal human-robot interaction to compose saliency-aware gaze heatmaps leveraged by a robot to learn emerging objects of interest. Our results show that our proposed method exceeds the capabilities of the current state of the art and outperforms it in terms of commonly used object detection metrics.

Autonomous vehicle navigation in complex and unpredictable outdoor environments requires extensive and detailed understanding of the surrounding area and compliance with the traffic rules. In this paper, we attempt to imitate human driver behavior towards autonomous navigation that is suitable for diverse, challenging environments, whether urban, semi-structured or rural-like. Our approach starts with a novel method that we propose for extracting free space area using RGB and LiDAR data, in combination with a rough topometric map for route planning. Local goals are extracted in the final drivable region, and the vehicle draws a local path in the free space that is approximately in line with the overall path via a lattice planner. Our method is evaluated both in the publicly available KITTI urban dataset and a custom-made dataset of a semi-structured environment. In both cases, the results highlight the potential of our approach for further advancements in autonomous navigation and the development of safer and more human-like behaviors in driverless vehicles compared to the existing trajectory prediction state-of-the-art methods that make use of a topometric map.

The safety of autonomous vehicles (AVs) depends on their ability to perform complex computations on high-volume sensor data in a timely manner. Their ability to run these computations with state-of-the-art models is limited by the processing power and slow update cycles of their onboard hardware. In contrast, cloud computing offers the ability to burst computation to vast amounts of the latest generation of hardware. However, accessing these cloud resources requires traversing wireless networks that are often considered to be too unreliable for real-time AV driving applications. Our work seeks to harness this unreliable cloud to enhance the accuracy of an AV's decisions, while ensuring that it can always fall back to its on-board computational capabilities. We identify three mechanisms that can be used by AVs to safely leverage the cloud for accuracy enhancements, and elaborate why current execution systems fail to enable these mechanisms. To address these limitations, we provide a system design based on the speculative execution of an AV's pipeline in the cloud, and show the efficacy of this approach in simulations of complex real-world scenarios that apply these mechanisms.

Does the norm of first-come-first-serve (FCFS) equally apply to those piloting a Mobile Remote Presence (MRP) system as to those who are physically present with it? While telepresence robots could make social interactions more accessible and enjoyable for geographically-constrained individuals, such an outcome requires both pilots and local users of MRPs to share the same social norm expectations that govern their use. To address this question, we conducted an online study (N=903)(N=903) involving simulated human-MRP interaction scenarios. Our results suggest that those remotely piloting the MRP-rather than local users-assign the robot to a lower social priority; they find it more appropriate when local users ignore queue order than when pilots ignore queue order. Furthermore, we provide significant empirical evidence that local users expect different social norms to be upheld depending on how they perceive the robot. Those who perceive MRPs simply as robots-rather than an extension of a person-do not expect the FCFS norm to be respected for MRPs.

Rescue robots require versatility and the capability to operate in various environments to carry out a diverse set of tasks effectively. The eccentric paddle (ePaddle) mechanism stands out for its high efficiency and adaptability. Generally, it is designed as a quadruped robot with a combined structure for fully-actuated control, this approach is often both inefficient and inflexible due to the requirement for repeated front-to-back paths. Unlike the fully-actuated controller that assume torso is fixed, this study proposes an under-actuated controller, consisting of a single ePaddle mechanism and a free torso for more efficient and flexible movement. Inspired by human gait, precision walking, and non-precision walking are introduced to discuss the stability of zero dynamics. Additionally, the stability condition is presented and demonstrated by numerical simulation. Since this control is based on robot dynamics, it has a high fault-tolerance and benefited from its dynamics attractor. The concept of under-actuated controller we proposed in this study is not only applicable to the ePaddle mechanism, but also to other under-actuated legged locomotion models.

When grasping objects with a multi-finger hand, it is crucial for the grasp stability to apply the correct torques at each joint so that external forces are countered. Most current systems use simple heuristics instead of modeling the required torque correctly. Instead, we propose a learning-based approach that is able to predict torques for grasps on unknown objects in real-time. The neural network, trained end-to-end using supervised learning, is shown to predict torques that are more efficient, and the objects are held with less involuntary movement compared to all tested heuristic baselines. Specifically, for 90 % of the grasps the translational deviation of the object is below 2.9 mm and the rotational below 3.1°. To generate training data, we formulate the analytical computation of torques as an optimization problem and handle the indeterminacy of multi-contacts using an elastic model. We further show that the network generalizes to predict torques for unknown objects on the real robot system with an inference time of 1.5 ms. Website: dlr-alr.github.io/grasping/

We consider the problem of time-limited robotic exploration in previously unseen environments where exploration is limited by a predefined amount of time. We propose a novel exploration approach using learning-augmented model-based planning. We generate a set of sub goals associated with frontiers on the current map and derive a Bellman Equation for exploration with these subgoals. Visual sensing and advances in semantic mapping of indoor scenes are exploited for training a deep convolutional neural network to estimate properties associated with each frontier: the expected unobserved area beyond the frontier and the expected time steps (discretized actions) required to explore it. The proposed model-based planner is guaranteed to explore the whole scene if time permits. We thoroughly evaluate our approach on a large-scale pseudo-realistic indoor dataset (Matterport3D) with the Habitat simulator. We compare our approach with classical and more recent RL-based exploration methods. Our approach surpasses the greedy strategies by 2.1% and the RL-based exploration methods by 8.4% in terms of coverage.

Demonstrations are widely used in Deep Reinforcement Learning (DRL) for facilitating solving tasks with sparse rewards. However, the tasks in real-world scenarios can often have varied initial conditions from the demonstration, which would require additional prior behaviours. For example, consider we are given the demonstration for the task of picking up an object from an open drawer, but the drawer is closed in the training. Without acquiring the prior behaviours of opening the drawer, the robot is unlikely to solve the task. To address this, in this paper we propose an Intrinsic Reward Driven Example-based Control (IRDEC). Our method can endow agents with the ability to explore and acquire the required prior behaviours and then connect to the task-specific behaviours in the demonstration to solve sparse-reward tasks without requiring additional demonstration of the prior behaviours. The performance of our method outperforms other baselines on three navigation tasks and one robotic manipulation task with sparse rewards. Codes are available at https://github.com/Ricky-Zhu/IRDEC.

In this work, we consider the problem of allocating tasks and planning paths for multiple robots to operate cooper-atively. We formulate the problem as a bi-level optimization that involves optimizing the scheduling of robots and the collision-free path for each robot. To address the complexity of the environment with obstacles, we introduce a congestion-aware state representation technique with the aid of graph neural networks. We also derive a joint scheduling policy using multi-agent reinforcement learning, and we propose an additional auxiliary loss that promotes coordination among the robots. Through empirical evaluation, we demonstrate the effectiveness of our approach in solving the combined task allocation and path planning problem in a cooperative multi-robot system.

The field of aerial manipulation has seen rapid advances, transitioning from push-and-slide tasks to interaction with articulated objects. The motion trajectory of these complex actions is usually hand-crafted or a result of online optimization methods like Model Predictive Control (MPC) or Model Predictive Path Integral (MPPI) control. However, these methods rely on heuristics or model simplifications to efficiently run on onboard hardware, limiting their robustness, and making them sensitive to disturbances and differences between the real environment and its model. In this work, we propose a Reinforcement Learning (RL) approach to learn reactive motion behaviors for a manipulation task while producing policies that are robust to disturbances and modeling errors. Specifically, we train a policy to perform a door-opening task with an Omnidirectional Micro Aerial Vehicle (OMAV). The policy is trained in a physics simulator and shown in the real world, where it is able to generalize also to door closing tasks never seen in training. We also compare our method against a state-of-the-art MPPI solution in simulation, showing a considerable increase in robustness and speed.

Using visual SLAM to map new environments requires time-consuming visits to all regions for data collection. We propose an approach to estimate maps of areas beyond the visible regions using a cheap and readily available modality of data-sound. We introduce the idea of an active audio-visual mapping agent. Besides collecting visual data, the proposed agent emits sounds during navigation, captures the echoes, and uses them to accurately map unknown areas. We propose a reinforcement learning based method that simultaneously trains models to 1) estimate a map from the visual data, 2) output navigation actions, 3) output the decision to emit a sound and 4) refine estimated mans using the cantured audio. Our agent is trained and tested on 85 real-world homes from the Matterport3D dataset using the Habitat and SoundSpaces simulators for visual and audio data. Our method, unlike visual-data reliant approaches, yields more accurate maps with broader environmental coverage. In addition, compared to an agent that continually emits sounds, we observe that intelligently choosing when to emit sounds leads to accurate maps obatined with greater efficiency.

When performing cloth-related tasks, such as garment hanging, it is often important to identify and grasp certain structural regions—a shirt's collar as opposed to its sleeve, for instance. However, due to cloth deformability, these manipulation activities, which are essential in domestic, health care, and industrial contexts, remain challenging for robots. In this paper, we focus on how to segment and grasp structural regions of clothes to enable manipulation tasks, using hanging tasks as case study. To this end, a neural network-based perception system is proposed to segment a shirt's collar from areas that represent the rest of the scene in a depth image. With a 10-minute video of a human manipulating shirts to train it, our perception system is capable of generalizing to other shirts regardless of texture as well as to other types of collared garments. A novel grasping strategy is then proposed based on the segmentation to determine grasping pose. Experiments demonstrate that our proposed grasping strategy achieves 92%, 80%, and 50% grasping success rates with one folded garment, one crumpled garment and three crumpled garments, respectively. Our grasping strategy performs considerably better than tested baselines that do not take into account the structural nature of the garments. With the proposed region segmentation and grasping strategy, challenging garment hanging tasks are successfully implemented using an open-loop control policy. Supplementary material is available at https://sites.google.com/view/garment-hanging