ICRA 2023 Papers

Evolutionary algorithms are commonly used to solve many complex optimization problems in such fields as robotics, industrial automation, and complex system design. Yet, their performance is limited when dealing with high-dimensional complex problems because they often require enormous computational resources to yield desired solutions, and they may easily trap into local optima. To solve this problem, this work proposes a Self-adaptive Teaching-learning-based Optimizer with an improved Radial basis function model and a sparse Autoencoder (STORA). In STORA, a Self-adaptive Teaching-learning-based Optimizer is designed to dynamically adjust parameters for balancing exploration and exploitation during its solution process. Then, a sparse autoencoder (SAE) is adopted as a dimension reduction method to compress search space into lower-dimensional one for more efficiently guiding population to converge towards global optima. Besides, an Improved Radial Basis Function model (IRBF) is designed as a surrogate model to balance training time and prediction accuracy. It is adopted to save computational resources for improving overall performance. In addition, a dynamic population allocation strategy is adopted to well integrate SAE and IRBF in STORA. We evaluate it by comparing it with several state-of-the-art algorithms through six benchmark functions. We further test it by applying it to solve a real-world computational offloading problem.

Depth estimation is critical for any robotic system. In the past years, the estimation of depth from monocular images has shown great improvement. However, in the underwater environment results are still lagging behind due to appearance changes caused by the medium. So far little effort has been invested in overcoming this. Moreover, underwater, there are more limitations to using high-resolution depth sensors, which is a serious obstacle to generating ground truth. So far unsupervised methods that tried to solve this have achieved limited success as they relied on domain transfer from a dataset in the air. We suggest network training using subsequent frames, self-supervised by a reprojection loss, as was demonstrated successfully above water. We propose several additions to the self-supervised framework to cope with the underwater environment and achieve state-of-the-art results on a challenging forward-looking underwater dataset.

This work proposes a self-supervised learning system for segmenting rigid objects in RGB images. The proposed pipeline is trained on unlabeled RGB-D videos of static objects, which can be captured with a camera carried by a mobile robot. A key feature of the self-supervised training process is a graph-matching algorithm that operates on the over-segmentation output of the point cloud that is reconstructed from each video. The graph matching, along with point cloud registration, is able to find reoccurring object patterns across videos and combine them into 3D object pseudo labels, even under occlusions or different viewing angles. Projected 2D object masks from 3D pseudo labels are used to train a pixel-wise feature extractor through contrastive learning. During online inference, a clustering method uses the learned features to cluster foreground pixels into object segments. Experiments highlight the method's effectiveness on both real and synthetic video datasets, which include cluttered scenes of tabletop objects. The proposed method outperforms existing unsupervised methods for object segmentation by a large margin.

When humans perform a task with an articulated object, they interact with the object only in a handful of ways, while the space of all possible interactions is nearly endless. This is because humans have prior knowledge about what interactions are likely to be successful, i.e., to open a new door we first try the handle. While learning such priors without supervision is easy for humans, it is notoriously hard for machines. In this work, we tackle unsupervised learning of priors of useful interactions with articulated objects, which we call interaction modes. In contrast to the prior art, we use no supervision or privileged information; we only assume access to the depth sensor in the simulator to learn the interaction modes. More precisely, we define a successful interaction as the one changing the visual environment substantially and learn a generative model of such interactions, that can be conditioned on the desired goal state of the object. In our experiments, we show that our model covers most of the human interaction modes, outperforms existing state-of-the-art methods for affordance learning, and can generalize to objects never seen during training. Additionally, we show promising results in the goal-conditional setup, where our model can be quickly fine-tuned to perform a given task. We show in the experiments that such affordance learning predicts interaction which covers most modes of interaction for the querying articulated object and can be fine-tuned to a goal-conditional model. For supplementary: https://actaim.github.io/.

In this work, we propose a self-improving artificial intelligence system to enhance the safety performance of reinforcement learning (RL)-based autonomous driving (AD) agents using black-box verification methods. RL algorithms have become popular in AD applications in recent years. However, the performance of existing RL algorithms heavily depends on the diversity of training scenarios. A lack of safety-critical scenarios during the training phase could result in poor generalization performance in real-world driving applications. We propose a novel framework in which the weaknesses of the training set are explored through black-box verification methods. After discovering AD failure scenarios, the RL agent's training is re-initiated via transfer learning to improve the performance of previously unsafe scenarios. Simulation results demonstrate that our approach efficiently discovers safety failures of action decisions in RL-based adaptive cruise control (ACC) applications and significantly reduces the number of vehicle collisions through iterative applications of our method. The source code is publicly available at https://github.com/data-and-decision-lab/self-improving-RL.

A novel mechanism to derive self-entanglement-free path for tethered differential-driven robots is proposed in this work. The problem is tailored to the applications of tethered robots without an omni-directional tether re-tractor which is often encountered when an omni-directional tether retracting mechanism is incapable of being jointly equipped with other geometrically complex devices (e.g. a manipulator), for instance the disaster recovery, spatial exploration, etc. Without a special consideration on the spatial relation between the pose of the mobile base and the tether, self-entanglement appears when the robot moves, resulting in unsafe motion of the robot and potential damage to the tether. In this paper, the self-entanglement-free constraint is modelled as the admissible orientation of the tether anchoring on the robot with respect to the robot's heading orientation. A searching-based path planning algorithm is then proposed to generate a near optimal path solution with guaranteed null of tether self-entanglement. The effectiveness of the proposed algorithm is compared with the motions without considering self-entanglement-free constraint, illustrated in challenging planning cases, and validated in realworld scenes. An open-source implementation has also been provided for the benefit of the robotics community.

Powered by deep representation learning, re-inforcement learning (RL) provides an end-to-end learning framework capable of solving self-driving (SD) tasks without manual designs. However, time-varying nonstationary environments cause proficient but specialized RL policies to fail at execution time. For example, an RL-based SD policy trained under sunny days does not generalize well to rainy weather. Even though meta learning enables the RL agent to adapt to new tasks/environments, its offline operation fails to equip the agent with online adaptation ability when facing nonstationary environments. This work proposes an online meta reinforcement learning algorithm based on the conjectural online lookahead adaptation (COLA). COLA determines the online adaptation at every step by maximizing the agent's conjecture of the future performance in a lookahead horizon. Experimental results demonstrate that under dynamically changing weather and lighting conditions, the COLA-based self-adaptive driving outperforms the baseline policies regarding online adaptability. A demo video, source code, and appendixes are available at https://github.com/Panshark/COLA

This paper presents Segregator, a global point cloud registration framework that exploits both semantic information and geometric distribution to efficiently build up outlier-robust correspondences and search for inliers. Current state-of-the-art algorithms rely on point features to set up putative correspondences and refine them by employing pair-wise distance consistency checks. However, such a scheme suffers from degenerate cases, where the descriptive capability of local point features downgrades, and unconstrained cases, where length-preserving (1-TRIMs)-based checks cannot sufficiently constrain whether the current observation is consistent with others, resulting in a complexified NP-complete problem to solve. To tackle these problems, on the one hand, we propose a novel degeneracy-robust and efficient corresponding procedure consisting of both instance-level semantic clusters and geometric-level point features. On the other hand, Gaussian distribution-based translation and rotation invariant measurements (G-TRIMs) are proposed to conduct the consistency check and further constrain the problem size. We validated our proposed algorithm on extensive real-world data-based experiments. The code is available: https://github.com/Pamphlett/Segregator.

In this paper, we investigate the problem of reinforcement learning under linear temporal logic (LTL) specifications for Markov decision processes (MDPs) with security constraints. We consider an outside passive intruder (observer) that can observe the external output behavior of the system through an output projection. We assume that the secret of the system is a subset of the initial states. The security constraint requires that the observer can never infer for sure that the agent was initiated from a secret state. Our objective is to learn a control policy that achieves the LTL task while ensuring security. To solve the problem of shaping the reward for reinforcement learning, we propose an approach based on the initial-state estimator and the limit deterministic Büchi automata. We illustrate the proposed approach by a case study of mobile robot example.

Multi-Agent Path Finding (MA-PF) computes a set of collision-free paths for multiple agents from their respective starting locations to destinations. This paper considers a generalization of MA-PF called Multi-Agent Teamwise Cooperative Path Finding (MA-TC-PF), where agents are grouped as multiple teams and each team has its own objective to be minimized. For example, an objective can be the sum or max of individual arrival times of the agents. In general, there is more than one team, and MA-TC-PF is thus a multi-objective planning problem with the goal of finding the entire Pareto-optimal front that represents all possible trade-offs among the objectives of the teams. To solve MA-TC-PF, we propose two algorithms TC-CBS and TC-M*, which leverage the existing CBS and M* for conventional MA-PF. We discuss the conditions under which the proposed algorithms are complete and are guaranteed to find the Pareto-optimal front. We present numerical results for several types of MA-TC-PF problems.

Accurate camera-to-lidar calibration is a requirement for sensor data fusion in many 3D perception tasks. In this paper, we present SceneCalib, a novel method for simultaneous self-calibration of extrinsic and intrinsic parameters in a system containing multiple cameras and a lidar sensor. Existing methods typically require specially designed calibration targets and human operators, or they only attempt to solve for a subset of calibration parameters. We resolve these issues with a fully automatic method that requires no explicit correspondences between camera images and lidar point clouds, allowing for robustness to many outdoor environments. Furthermore, the full system is jointly calibrated with explicit cross-camera constraints to ensure that camera-to-camera and camera-to-lidar extrinsic parameters are consistent.

In robotic applications, we often obtain tons of 3D point cloud data without color information, and it is difficult to visualize point clouds in a meaningful and colorful way. Can we colorize 3D point clouds for better visualization? Existing deep learning-based colorization methods usually only take simple 3D objects as input, and their performance for complex scenes with multiple objects is limited. To this end, this paper proposes a novel semantics-and-geometry-aware colorization network, termed SGNet, for vivid scene-level point cloud colorization. Specifically, we propose a novel pipeline that explores geometric and semantic cues from point clouds containing only coordinates for color prediction. We also design two novel losses, including a colorfulness metric loss and a pairwise consistency loss, to constrain model training for genuine colorization. To the best of our knowledge, our work is the first to generate realistic colors for point clouds of large-scale indoor scenes. Extensive experiments on the widely used ScanNet benchmarks demonstrate that the proposed method achieves state-of-the-art performance on point cloud colorization.

In recent years, reinforcement learning and its multi-agent analogue have achieved great success in solving various complex control problems. However, multi-agent rein-forcement learning remains challenging both in its theoretical analysis and empirical design of algorithms, especially for large swarms of embodied robotic agents where a definitive toolchain remains part of active research. We use emerging state-of-the-art mean-field control techniques in order to convert many-agent swarm control into more classical single-agent control of distributions. This allows profiting from advances in single-agent reinforcement learning at the cost of assuming weak interaction between agents. However, the mean-field model is violated by the nature of real systems with embodied, physically colliding agents. Thus, we combine collision avoidance and learning of mean-field control into a unified framework for tractably designing intelligent robotic swarm behavior. On the theoretical side, we provide novel approximation guarantees for general mean-field control both in continuous spaces and with collision avoidance. On the practical side, we show that our approach outperforms multi-agent reinforcement learning and allows for decentralized open-loop application while avoiding collisions, both in simulation and real UAV swarms. Overall, we propose a framework for the design of swarm behavior that is both mathematically well-founded and practically useful, enabling the solution of otherwise intractable swarm problems.

Existing spatial localization techniques for au-tonomous vehicles mostly use a pre-built 3D-HD map, often constructed using a survey-grade 3D mapping vehicle, which is not only expensive but also laborious. This paper shows that by using an off-the-shelf high-definition satellite image as a ready-to-use map, we are able to achieve cross-view vehicle localization up to a satisfactory accuracy, providing a cheaper and more practical way for localization. While the utilization of satellite imagery for cross-view localization is an established concept, the conventional methodology focuses primarily on image re-trieval. This paper introduces a novel approach to cross-view localization that departs from the conventional image retrieval method. Specifically, our method develops (1) a Geometric-align Feature Extractor (GaFE) that leverages measured 3D points to bridge the geometric gap between ground and overhead views, (2) a Pose Aware Branch (PAB) adopting a triplet loss to encourage pose-aware feature extraction, and (3) a Recursive Pose Refine Branch (RPRB) using the Levenberg-Marquardt (LM) algorithm to align the initial pose towards the true vehicle pose iteratively. Our method is validated on KITTI and Ford Multi-AV Seasonal datasets as ground view and Google Maps as the satellite view. The results demonstrate the superiority of our method in cross-view localization with median spatial and angular errors within 1 meter and 1°, respectively.

Replanning in temporal logic tasks is extremely difficult during the online execution of robots. This study introduces an effective path planner that computes solutions for temporal logic goals and instantly adapts to non-static and partially unknown environments. Given prior knowledge and a task specification, the planner first identifies an initial feasible solution by growing a sampling-based search tree. While carrying out the computed plan, the robot maintains a solution library to continuously enhance the unfinished part of the plan and store backup plans. The planner updates existing plans when meeting unexpected obstacles or recognizing flaws in prior knowledge. Upon a high-level path is obtained, a trajectory generator tracks the path by dividing it into segments of motion primitives. Our planner is integrated into an autonomous mobile robot system, further deployed on a multicopter with limited onboard processing power. In simulation and real-world experiments, our planner is demonstrated to swiftly and effectively adjust to environmental uncertainties.

We propose Predictive Information bottleneck for Goal representation learning (PI-Goal), a self-supervised method for sample-efficient goal-conditioned reinforcement learning (RL). Goal-conditioned RL learns to reach commanded goals with reward signals. A goal could be given in a noisy or abstract form, and thus jeopardizes sample efficiency. Previous methods usually assume that the agent can map a state to an achievable goal. In this work, we consider a setting in which the goal space is unknown to the agent and the agent cannot recognize a goal in a specific state (referred to as a goal state) until the goal is commanded. Our PI-Goal learns a goal representation which contains only the predictive information of a goal state, i.e., the mutual information between a current state and a future state, and guarantees the optimality of the learned policy. Experimental results show that PI-Goal consistently outperforms the baseline methods in tasks with unknown goal spaces, e.g., object manipulation, object search, and embodied question answering.

Sampling-based motion planning works well in many cases but is less effective if the configuration space has narrow passages. In this paper, we propose a learning-based strategy to sample in these narrow passages, which improves overall planning time. Our algorithm first learns from the configuration space planning graphs and then uses the learned information to effectively generate narrow passage samples. We perform experiments in various 6D and 7D scenes. The algorithm offers one order of magnitude speed-up compared to baseline planners in some of these scenes.

Deep learning has led to great progress in the detection of mobile (i.e. movement-capable) objects in urban driving scenes in recent years. Supervised approaches typically require the annotation of large training sets; there has thus been great interest in leveraging weakly, semi- or self- supervised methods to avoid this, with much success. Whilst weakly and semi-supervised methods require some annotation, self-supervised methods have used cues such as motion to relieve the need for annotation altogether. However, a complete absence of annotation typically degrades their performance, and ambiguities that arise during motion grouping can inhibit their ability to find accurate object boundaries. In this paper, we propose a new self-supervised mobile object detection approach called SCT. This uses both motion cues and expected object sizes to improve detection performance, and predicts a dense grid of 3DD oriented bounding boxes to improve object discovery. We significantly outperform the state-of-the-art self-supervised mobile object detection method TCR on the KITTI tracking benchmark, and achieve performance that is within 30 % of the fully supervised PV-RCNN++ method for IoUs \leq\leq 0.5. Our source code will be made available online.

Model generalization of the underlying dynamics is critical for achieving data efficiency when learning for robot control. This paper proposes a novel approach for learning dynamics leveraging the symmetry in the underlying robotic system, which allows for robust extrapolation from fewer samples. Existing frameworks that represent all data in vector space fail to consider the structured information of the robot, such as leg symmetry, rotational symmetry, and physics invariance. As a result, these schemes require vast amounts of training data to learn the system's redundant elements because they are learned independently. Instead, we propose considering the geometric prior by representing the system in symmetrical object groups and designing neural network architecture to assess invariance and equivariance between the objects. Finally, we demonstrate the effectiveness of our approach by comparing the generalization to unseen data of the proposed model and the existing models. We also implement a controller of a climbing robot based on learned inverse dynamics models. The results show that our method generates accurate control inputs that help the robot reach the desired state while requiring less training data than existing methods.