IROS 2024 Papers

Reinforcement Learning (RL) algorithms can learn robotic control tasks from visual observations, but they often require a large amount of data, especially when the visual scene is complex and unstructured. In this paper, we explore how the agent’s knowledge of its shape can improve the sample efficiency of visual RL methods. We propose a novel method, Disentangled Environment and Agent Representations (DEAR), that uses the segmentation mask of the agent as supervision to learn disentangled representations of the environment and the agent through feature separation constraints. Unlike previous approaches, DEAR does not require reconstruction of visual observations. These representations are then used as an auxiliary loss to the RL objective, encouraging the agent to focus on the relevant features of the environment. We evaluate DEAR on two challenging benchmarks: Distracting DeepMind control suite and Franka Kitchen manipulation tasks. Our findings demonstrate that DEAR surpasses state-of-the-art methods in sample efficiency, achieving comparable or superior performance with reduced parameters. Our results indicate that integrating agent knowledge into visual RL methods has the potential to enhance their learning efficiency and robustness.

Estimating camera motion and continuously reconstructing dense scenes in deformable environments presents a complex and open challenge. Many existing approaches tend to rely on assumptions about the scene’s topology or the nature of deformable motion. However, these assumptions do not hold true in medical endoscopy applications. To address these challenges, we introduce DDS-SLAM, a novel dense deformable semantic neural SLAM that achieves accurate camera tracking, continuous dense scene reconstruction, and high-quality image rendering in deformable scenes. First, we propose a novel hybrid neural scene representation method capable of capturing both natural and artificial deformations. Additionally, by leveraging the 2D semantic information of the scene, we introduce a semantic loss function based on semantic distance fields. This approach guides network optimization at a higher level, thereby enhancing system performance. Furthermore, we validate our method through a series of experiments conducted on several representative medical datasets, demonstrating its superiority over other state-of-the-art approaches. The code is available at: https://github.com/IRMVLab/DDS-SLAM.

Multimodal image pairs can make object detection more reliable in challenging environments, so RGB-T object detection has gained extensive attention over the past decade. To alleviate the complementarity of the visible and thermal modality, we propose a novel lightweight Feature Enhancement-fusion Module (FEM), which is composed of the Channel Enhancement-fusion Unit (CEU) and Spatial Enhancement-fusion Unit (SEU) by extending the attention mechanism to operate on two modalities. CEU is used to exploit the complementarity and alleviate the data imbalance by combining internal and global channel attention. Additionally, SEU is utilized to guide the model to pay more attention to the regions of interest. By incorporating FEM, enhanced and fused features are obtained, leading to improved performance. The effectiveness and generalizability of FEM are validated by two public datasets and our proposed DCSANet achieves competitive performance while maintaining high speed (+%7.0 on LLVIP and +1.2% on FLIR in mAP). Moreover, we conducted ablation experiments to verify the effectiveness of the proposed operators.

Solving storage problems—where objects must be accurately placed into containers with precise orientations and positions—presents a distinct challenge that extends beyond traditional rearrangement tasks. These challenges are primarily due to the need for fine-grained 6D manipulation and the inherent multi-modality of solution spaces, where multiple viable goal configurations exist for the same storage container. We present a novel Diffusion-based Affordance Prediction (DAP) pipeline for the multi-modal object storage problem. DAP leverages a two-step approach, initially identifying a placeable region on the container and then precisely computing the relative pose between the object and that region. Existing methods either struggle with multi-modality issues or computation-intensive training. Our experiments demonstrate DAP’s superior performance and training efficiency over the current state-of-the-art RPDiff, achieving remarkable results on the RPDiff benchmark. Additionally, our experiments showcase DAP’s data efficiency in real-world applications, an advancement over existing simulation-driven approaches. Our contribution fills a gap in robotic manipulation research by offering a solution that is both computationally efficient and capable of handling real-world variability. Code and supplementary material can be found at: https://github.com/changhaonan/DPS.git.

This paper develops a new Distributed approach for solving the inverse problem of a Differentiable Dynamic Game (D3G), which enables robots to learn multi-robot coordination from given demonstrations. We formulate multi-robot coordination as the Nash equilibrium of a parameterized dynamic game, where the behavior of each robot is dictated by an objective function that also depends on the behavior of its neighboring robots. The coordination thus can be adapted by tuning the parameters of the objective and the local dynamics of each robot. The proposed algorithm enables each robot to automatically tune such parameters in a distributed and coordinated fashion — only using the data of its neighbors without global information. Its key novelty is the development of a distributed solver for a diff-KKT condition that can enhance scalability and reduce the computational load for gradient computation. We test the proposed algorithm in simulation with heterogeneous robots given different task configurations. The results demonstrate its effectiveness and generalizability for learning multi-robot coordination from demonstrations.

We address the challenge of multi-LiDAR interference, an issue of growing importance as LiDAR sensors are embedded in a growing set of pervasive devices. We introduce a novel approach named D2SR, enabling decentralized interference detection, mitigation, and recovery without explicit coordination among nearby LiDAR devices. D2SR comprises three stages: (a) Detection, which identifies interfered frames, (b) Mitigation, which performs time-shifting of a LiDAR’s active period to reduce interference, and (c) Recovery, which corrects or reconstructs the depth values in interfered regions of a depth frame. Key contributions include a lightweight interference detection algorithm achieving an F1-score of 92%, a simple yet effective decentralized de-synchronization mechanism, and a lightweight depth recovery pipeline that preserves high throughput processing on edge devices. Evaluation on Nvidia Jetson devices demonstrates D2SR’s efficacy: under static settings, D2SR accurately detects interference in 93% of cases (recall=82%) and reduces the depth estimation error by 27% (RMSE= 38.7 cm, compared to RMSE= 60.6 cm for a baseline without D2SR). Furthermore, D2SR is able to reduce the fraction of interfered frames by 75.1% and reduce the depth estimation error (for interfered frames) by 24.9% even for a moving robot scenario.

In this article, we propose a novel communication-based action space enhancement for the D-MARL exploration algorithm to improve the efficiency of mapping an unknown environment, represented by an occupancy grid map. In general, communication between autonomous systems is crucial when exploring large and unstructured environments. In such real-world scenarios, data transmission is limited and relies heavily on inter-agent proximity and the attributes of the autonomous platforms. In the proposed approach, each agent’s policy is optimized by utilizing the heterogeneous-agent proximal policy optimization algorithm to autonomously choose whether to communicate or explore the environment. To accomplish this, multiple novel reward functions are formulated by integrating inter-agent communication and exploration. The investigated approach aims to increase efficiency and robustness in the mapping process, minimize exploration overlap, and prevent agent collisions. The D-MARL policies trained on different reward functions have been compared to understand the effect of different reward terms on the collaborative attitude of the homogeneous agents. Finally, multiple simulation results are provided to prove the efficacy of the proposed scheme.

Visual Place Recognition (VPR) techniques commonly utilize Contrastive Losses (CL) to train models that generate compact and discriminative global descriptors for images. These models often result in poor performance due to one of the following reasons during training: 1) loss functions that focus primarily on easier samples, 2) reliance on time-consuming hard sample mining methods to identify informative supervisory samples, which hinders effective learning from large-scale datasets. To enhance both learning efficiency and effectiveness, we propose a Curricular Contrastive Loss (CCL) and use graded similarity labels as a measure of sample difficulty. Inspired by human learning that begin with easier concepts and progressively tackle more challenging ones, our CCL dynamically emphasizes easier samples during the initial training stages to achieve rapid convergence. The learning gradually focuses on harder samples in later training stages to bolster robustness of the models under challenging conditions. Our proposed method has been extensively evaluated on popular datasets, and the results demonstrate its superior performance compared to the CL and Generalized CL functions.

As robots shift from industrial to human-centered spaces, adopting mobile manipulators, which expand workspace capabilities, becomes crucial. In these settings, seamless interaction with humans necessitates compliant control. Two common methods for safe interaction, admittance, and impedance control, require force or torque sensors, often absent in lower-cost or lightweight robots. This paper presents an adaption of impedance control that can be used on current-controlled robots without the use of force or torque sensors and shows its application for compliant control of a mobile manipulator. A calibration method is designed that enables estimation of the actuators’ current/torque ratios and frictions, used by the adapted impedance controller, and that can handle model errors. The calibration method and the performance of the designed controller are experimentally validated using the Kinova GEN3 Lite arm. Results show that the calibration method is consistent and that the designed controller for the arm is compliant while also being able to track targets with five-millimeter precision when no interaction is present. Additionally, this paper presents two operational modes for interacting with the mobile manipulator: one for guiding the robot around the workspace through interacting with the arm and another for executing a tracking task, both maintaining compliance to external forces. These operational modes were tested in real-world experiments, affirming their practical applicability and effectiveness.Code: https://github.com/tud-amr/mobile-manipulator-compliance

A wire-driven parallel robot is a type of robotic system where multiple wires are used to control the movement of a end-effector. The wires are attached to the end-effector and anchored to fixed points on external structures. This configuration allows for the separation of actuators and end-effectors, enabling lightweight and simplified movable parts in the robot. However, its range of motion remains confined within the space formed by the wires, limiting the wire-driven capability to only within the pre-designed operational range. Here, in this study, we develop a wire-driven robot, CubiX, capable of connecting to and utilizing the environment. CubiX connects itself to the environment using up to 8 wires and drives itself by winding these wires. By integrating actuators for winding the wires into CubiX, a portable wire-driven parallel robot is realized without limitations on its workspace. Consequently, the robot can form parallel wire-driven structures by connecting wires to the environment at any operational location.

Mobile robot navigation in crowded environments with pedestrians is a crucial challenge in realizing service robots that can assist people in their daily lives. Navigation methods for mobile robots in environments employing deep reinforcement learning have been extensively studied. However, addressing such unexpected situations is a significant challenge. This study presents an approach that discerns whether a situation has been supposed to utilize a normalizing flow and dynamically switches between learning- and rule-based methods. Specifically, the proposed method achieves a higher success rate than employing only a learning-based approach and reaches the destination faster than employing only a rule-based approach in unexpected situations. Experiments are conducted to validate the performance enhancement achieved with the proposed switching method in both simulated and real-world settings.

This study seeks towards a better understanding of multi-vehicle routing problems (VRPs) under restricted observability. Unlike most prior research that assumes full knowledge of tasks and vehicles, this paper addresses VRPs where each vehicle’s observation is confined to the k-nearest neighbourhood. Vehicles make decisions based on localized policies in a decentralized manner. We theoretically demonstrate that for the imitation policy, the upper bound of the optimality gap diminishes as the neighbourhood range expands. Subsequently, we employed a multi-agent cross-observability policy optimization (MACOPO) algorithm to solve the VRPs with restricted observability. The algorithm optimizes a cross-entropy term by leveraging a fully observable expert to guide the training. Empirical results supported both the theoretical findings and the effectiveness of the multi-agent learning algorithm.

Cross-modal localization has drawn increasing attention in recent years, while the visual relocalization in prior LiDAR maps is less studied. Related methods usually suffer from inconsistency between the 2D texture and 3D geometry, neglecting the intensity features in the LiDAR point cloud. In this paper, we propose a cross-modal visual relocalization system in prior LiDAR maps utilizing intensity textures, which consists of three main modules: map projection, coarse retrieval, and fine relocalization. In the map projection module, we construct the database of intensity channel map images leveraging the dense characteristic of panoramic projection. The coarse retrieval module retrieves the top-K most similar map images to the query image from the database, and retains the top-K’ results by covisibility clustering. The fine relocalization module applies a two-stage 2D-3D association and a covisibility inlier selection method to obtain robust correspondences for 6DoF pose estimation. The experimental results on our self-collected datasets demonstrate the effectiveness in both place recognition and pose estimation tasks.

3D perception in LiDAR point clouds is crucial for a self-driving vehicle to properly act in 3D environment. However, manually labeling point clouds is hard and costly. There has been a growing interest in self-supervised pre-training of 3D perception models. Following the success of contrastive learning in images, current methods mostly conduct contrastive pre-training on point clouds only. Yet an autonomous driving vehicle is typically supplied with multiple sensors including cameras and LiDAR. In this context, we systematically study single modality, cross-modality, and multi-modality for contrastive learning of point clouds, and show that cross-modality wins over other alternatives. In addition, considering the huge difference between the training sources in 2D images and 3D point clouds, it remains unclear how to design more effective contrastive units for LiDAR. We therefore propose the instance-aware and similarity-balanced contrastive units that are tailored for self-driving point clouds. Extensive experiments reveal that our approach achieves remarkable performance gains over various point cloud models across the downstream perception tasks of LiDAR based 3D object detection and 3D semantic segmentation on the four popular benchmarks including Waymo Open Dataset, nuScenes, SemanticKITTI and ONCE.

Recent years have seen object detection robotic systems deployed in several personal devices (e.g., home robots and appliances). This has highlighted a challenge in their design, i.e., they cannot efficiently update their knowledge to distinguish between general classes and user-specific instances (e.g., a dog vs. user’s dog). We refer to this challenging task as Instance-level Personalized Object Detection (IPOD). The personalization task requires many samples for model tuning and optimization in a centralized server, raising privacy concerns. An alternative is provided by approaches based on recent large-scale Foundation Models, but their compute costs preclude on-device applications. In our work we tackle both problems at the same time, designing a Few-Shot IPOD strategy called AuXFT. We introduce a conditional coarse-to-fine few-shot learner to refine the coarse predictions made by an efficient object detector, showing that using an off-the-shelf model leads to poor personalization due to neural collapse. Therefore, we introduce a Translator block that generates an auxiliary feature space where features generated by a self-supervised model (e.g., DINOv2) are distilled without impacting the performance of the detector. We validate AuXFT on three publicly available datasets and one in-house benchmark designed for the IPOD task, achieving remarkable gains in all considered scenarios with excellent time-complexity trade-off: AuXFT reaches a performance of 80% its upper bound at just 32% of the inference time, 13% of VRAM and 19% of the model size.

In this work, we propose a method of capturing the patient’s discomfort during robotic shoulder physiotherapy, creating "discomfort maps". These maps depict the personalized distribution of discomfort that each patient perceived across their shoulder range of motion, facilitating both robotic devices and human therapists to account for patient-specific characteristics during the therapeutic process. Our system enables a patient to communicate and map discomfort in space and time during movement via a handheld push-button device, while interacting with a robotic physical therapy device capable of moving the patient and estimating their pose. We validated our method through human factors experiments simulating shoulder physiotherapy sessions with 10 healthy participants. To avoid the risk of injury to the participants and to allow for ground truth map information, we emulate perceived discomfort via an auditory signal. Our experimental apparatus enabled participants to reconstruct synthetic discomfort maps, demonstrating the feasibility of automatically capturing and storing patient discomfort during robotic physiotherapy.

Recently, Shuttle-based Storage and Retrieval Systems (SBS/RSs) have garnered significant attention from both academia and industry, owing to their high spatial utilization and rapid response speed. However, the weak connectivity of roadmaps in densely stored environments increases the likelihood of congestion and deadlocks when multiple four-way shuttles operate simultaneously, thereby imposing greater demands for collaborative path planning. Instead of exhaustively coordinating the shuttle motions during the off-line planning or online local control stages, we solve the cooperative path planning challenge from the perspective of altering the road graph structure dynamically. More specifically, we propose an approach to automatically transfer the typical road graph of SBS/RSs into a hierarchical graph with a reduced size, and then dynamically adjust its edge properties to prohibit any motion conflicts. In this manner, the planning problem of large-scale shuttle groups can be easily resolved and all the potential congestions can be eliminated inherently. Finally, we build a complete multi-shuttle cooperative path planning system adaptable for large-scale problems.

Soldiers in the field often need to cross negative obstacles, such as rivers or canyons, to reach goals or safety. Military gap crossing involves on-site temporary bridges construction. However, this procedure is conducted with dangerous, time and labor intensive operations, and specialized machinery. We envision a scalable robotic solution inspired by advancements in force-controlled and Cable-Driven Parallel Robots (CDPRs); this solution can address the challenges inherent in this transportation problem, achieving fast, efficient, and safe deployment and field operations. We introduce the embodied vision in Co3MaNDR, a solution to the military gap crossing problem, a distributed robot consisting of several modules simultaneously pulling on a central payload, controlling the cables’ tensions to achieve complex objectives, such as precise trajectory tracking or force amplification. Hardware experiments demonstrate teleoperation of a payload, trajectory following, and the sensing and amplification of operators’ applied physical forces during slow operations. An operator was shown to manipulate a 27.2 kg (60 lb) payload with an average force utilization of 14.5% of its weight. Results indicate that the system can be scaled up to heavier payloads without compromising performance or introducing superfluous complexity. This research lays a foundation to expand CDPR technology to uncoordinated and unstable mobile platforms in unknown environments.

Hysteresis has posed significant challenges to the modeling and control of flexible endoscopic robots, which impedes the advancement of automated endoscopic operation. Despite numerous hysteresis modeling approaches aimed at improving accuracy, there are still several unresolved issues, such as inappropriate model selection and non-ideal assumption of noise. Focusing on these challenges, a novel reinforcement active modeling (RAM) scheme is proposed in this paper. By incorporating reinforcement learning, this method augments an Extended Kalman Filter (EKF)-based active modeling strategy, which improves the insensitivity and generalization ability to non-Gaussian noise that is not introduced in training. Finally, a series of comparative experiments are conducted on the self-built flexible endoscopic robot to validate the improvement achieved by the proposed scheme. Compared with some widely-applied methods, the proposed scheme achieved at least 63.8% improvement in the root mean square error (RMSE) in modeling accuracy under Gaussian noise conditions, and at least 36.5% improvement in RMSE under Poisson noise conditions.