ICRA 2023 论文列表

Autonomous Underwater Vehicles (AUVs) conduct regular visual surveys of marine environments to characterise and monitor the composition and diversity of the benthos. The use of machine learning classifiers for this task is limited by the low numbers of annotations available and the many fine-grained classes involved. In addition to these challenges, there are domain shifts between image sets acquired during different AUV surveys due to changes in camera systems, imaging altitude, illumination and water column properties leading to a drop in classification performance for images from a different survey where some or all these elements may have changed. This paper proposes a framework to improve the performance of a benthic morphospecies classifier when used to classify images from a different survey compared to the training data. We adapt the SymmNet state-of-the-art Unsupervised Domain Adaptation method with an efficient bilinear pooling layer and image scaling to normalise spatial resolution, and show improved classification accuracy. We test our approach on two datasets with images from AUV surveys with different imaging payloads and locations. The results show that generic domain adaptation can be enhanced to produce a significant increase in accuracy for images from an AUV survey that differs from the training images.

Teleoperated techniques enable remote human-robot interaction and have been widely accepted in robot-assisted surgeries. However, it is still hard to guarantee the safety of teleoperated surgery due to the imperfect input commands limited by remote perception, preventing teleoperated surgery from being widely used. We propose a new framework to avoid the collision of surgery robots and human tissue caused by inaccurate inputs. We directly take the medical volume data and propose to use the implicit neural field to guide teleoperated robot-assisted surgery. With guidance, the trajectory of the robot manipulator is optimized to safely work inside a narrow workspace. We evaluated our method in several aspects and conducted a real-world experiment on a head phantom. Experimental results show that our proposed method can effectively avoid the collision between the surgical tool and the human tissue during teleoperation.

Robust and efficient grasping of different objects is still an open problem due to the difficulty of integrating multidisciplinary knowledge such as gripper ontology design, perception, control, and learning. In recent years, learning-based methods have achieved excellent results in grasping various novel objects. However, current methods are usually limited to a single grasping mode or rely on different end effectors to grasp objects of different shapes. For human beings, our hands are capable of grasping various objects with changes in grasping methods and form of hands. In light of this, developing a gripper with similar performance could possibly improve the robot's gripping ability. In this paper, we design a dual-modal soft gripper (DSG) and propose a deep reinforcement learning (DRL) framework to implement the operations. Both of our grasping modes, namely enveloping and pinching, are achieved through the tendon drive system and the deformation of the spring steel plate, which enables the gripper to switch between the two grasping modes in real time. We also combined the cutting-edge achievements of deep learning and reinforcement learning to design an autonomous grasping algorithm based on Q-learning and a deep Q network. Moreover, to fully utilize the visual input from the sensor, we added semantic embeddings of target objects to facilitate the learning, which is especially useful in deciding the grasping method for objects previously unseen. We also evaluate our DRL framework in different scenarios, offering a detailed comparison of each grasping mode and the mixed method (with or without semantic information). Our design has proved efficient in reducing the number of failing grasping actions and improving the success rate when facing novel and tricky objects.

Achieving successful robotic manipulation is an essential step towards robots being widely used in industry and home settings. Recently, many learning-based methods have been proposed to tackle this challenge, with imitation learning showing great promise. However, imperfect demonstrations and a lack of feedback from teleoperation systems may lead to poor or even unsafe results. In this work we explore the effect of demonstrator force feedback on imitation learning, using a feedback glove and a robot arm to render fingertip-level and palm-level forces, respectively. 10 participants recorded 5 demonstrations of a pick-and-place task with 3 grippers, under conditions with no force feedback, fingertip force feedback, and fingertip and palm force feedback. Results show that force feedback significantly reduces demonstrator fingertip and palm forces, leads to a lower variation in demonstrator forces, and recorded trajectories that are quicker to execute. Using behavioral cloning, we find that agents trained to imitate these trajectories mirror these benefits, even though agents have no force data shown to them during training. We conclude that immersive demonstrations, achieved with force feedback, may be the key to unlocking safer, quicker-to-execute dexterous manipulation policies.

3D human reconstruction from RGB images achieves decent results in good weather conditions but degrades dramatically in rough weather. Complementary, mmWave radars have been employed to reconstruct 3D human joints and meshes in rough weather. However, combining RGB and mmWave signals for robust all-weather 3D human reconstruction is still an open challenge, given the sparse nature of mmWave and the vulnerability of RGB images. In this paper, we present ImmFusion, the first mmWave-RGB fusion solution to reconstruct 3D human bodies in all weather conditions robustly. Specifically, our ImmFusion consists of image and point backbones for token feature extraction and a Transformer module for token fusion. The image and point backbones refine global and local features from original data, and the Fusion Transformer Module aims for effective information fusion of two modalities by dynamically selecting informative tokens. Extensive experiments on a large-scale dataset, mmBody, captured in various environments demonstrate that ImmFusion can efficiently utilize the information of two modalities to achieve a robust 3D human body reconstruction in all weather conditions. In addition, our method's accuracy is significantly superior to that of state-of-the-art Transformer-based LiDAR-camera fusion methods.

A self-driving car must be able to reliably handle adverse weather conditions (e.g., snowy) to operate safely. In this paper, we investigate the idea of turning sensor inputs (i.e., images) captured in an adverse condition into a benign one (i.e., sunny), upon which the downstream tasks (e.g., semantic segmentation) can attain high accuracy. Prior work primarily formulates this as an unpaired image-to-image translation problem due to the lack of paired images captured under the exact same camera poses and semantic layouts. While perfectly-aligned images are not available, one can easily obtain coarsely-paired images. For instance, many people drive the same routes daily in both good and adverse weather; thus, images captured at close-by GPS locations can form a pair. Though data from repeated traversals are unlikely to capture the same foreground objects, we posit that they provide rich contextual information to supervise the image translation model. To this end, we propose a novel training objective leveraging coarsely-aligned image pairs. We show that our coarsely-aligned training scheme leads to a better image translation quality and improved downstream tasks, such as semantic segmentation, monocular depth estimation, and visual localization.

State estimation from measured data is crucial for robotic applications as autonomous systems rely on sensors to capture the motion and localize in the 3D world. Among sensors that are designed for measuring a robot's pose, or for soft robots, their shape, vision sensors are favorable because they are information-rich, easy to set up, and cost-effective. With recent advancements in computer vision, deep learning-based methods no longer require markers for identifying feature points on the robot. However, learning-based methods are data-hungry and hence not suitable for soft and prototyping robots, as building such bench-marking datasets is usually infeasible. In this work, we achieve image-based robot pose estimation and shape reconstruction from camera images. Our method requires no precise robot meshes, but rather utilizes a differentiable renderer and primitive shapes. It hence can be applied to robots for which CAD models might not be available or are crude. Our parameter estimation pipeline is fully differentiable. The robot shape and pose are estimated iteratively by back-propagating the image loss to update the parameters. We demonstrate that our method of using geometrical shape primitives can achieve high accuracy in shape reconstruction for a soft continuum robot and pose estimation for a robot manipulator.

Confined and cluttered aquatic environments present a number of significant challenges with respect to inspection by robotic platforms, including localisation and communications. Some of these can be mitigated by using collaborative heterogeneous multi-robot teams. An important element of such a system is collaborative control. This paper addresses this challenge by presenting an Image-Based Visual Servoing (IBVS), leader-follower control system for heterogeneous aquatic robots. Experiments were conducted in an uncluttered pond to demonstrate the capabilities of the system. The results show robots can maintain tracking each other with maximum xx and yy displacements of 0.42 m and 0.41 m, the maximum projection distance in the xy-plane of maintaining formation is 0.45 m, showing the stability and feasibility of deploying such system on underwater platforms.

In this work, we address the problem of robust segmentation of a continuum robot from images without the need for training data or markers. We present a method that leverages information about the kinematics of these robots to produce an estimate of the robot shape, which is refined through optimization over global image statistics. Our approach can be straightforwardly applied to any continuum robot design and is able to handle partial occlusions of the robot body, as well as challenging background conditions. We validate our method experimentally for a concentric tube robot in a simulated surgical environment and show that our method significantly outperforms a naive projection of the robot shape and color thresholding, which is commonly used in current vision-based estimation algorithms for these robots. Overall, this work has the potential to improve the viability of vision-based state estimation for continuum robots in real-world settings.

Self-supervised monocular depth estimation is a salient task for 3D scene understanding. Learned jointly with monocular ego-motion estimation, several methods have been proposed to predict accurate pixel-wise depth without using labeled data. Nevertheless, these methods focus on improving performance under ideal conditions without natural or digital corruptions. The general absence of occlusions is assumed even for object-specific depth estimation. These methods are also vulnerable to adversarial attacks, which is a pertinent concern for their reliable deployment in robots and autonomous driving systems. We propose MIMDepth, a method that adapts masked image modeling (MIM) for self-supervised monocular depth estimation. While MIM has been used to learn generalizable features during pre-training, we show how it could be adapted for direct training of monocular depth estimation. Our experiments show that MIMDepth is more robust to noise, blur, weather conditions, digital artifacts, occlusions, as well as untargeted and targeted adversarial attacks.

Whisker-based tactile sensors have the potential to perform fast and accurate 3D mappings of the environment, complementing vision-based methods under conditions of glare, reflection, proximity, and occlusion. However, current algorithms for mapping with whiskers make assumptions about the conditions of contact, and these assumptions are not always valid and can cause significant sensing errors. Here we introduce a new whisker sensing system with a tapered, flexible whisker. The system provides inputs to two separate algorithms for estimating radial contact distance on a whisker. Using a Gradient-Moment (GM) algorithm, we correctly detect contact distance in most cases (within 4% of the whisker length). We introduce the Z-Dissimilarity score as a new metric that quantifies uncertainty in the radial contact distance estimate using both the GM algorithm and a Moment-Force (MF) algorithm that exploits the tapered whisker design. Combining the two algorithms ultimately results in contact distance estimates more robust than either algorithm alone.

Identifying the inertial parameters of real robotic manipulators is a fundamental step towards realistic modeling and better controller performances, which is crucial for safe human-robot interaction. Our work introduces a novel framework for identifying a generalized set of base inertial parameters of a serial link manipulator. This framework is designed to be adaptable to accommodate any new mounting configuration of the robot. Our theoretical analysis highlights the influence of the robot's mounting configuration on the emergence of new parameters that cannot be identified through the conventional vertical base-axis mounting approach studied previously. To validate our proposed framework, we carried out two main experiments: the first involved simulation to establish the feasibility of our concept, and in the second, our framework was employed on a Franka Emika Robot in a real-world scenario to demonstrate and validate our approach. Our simulation results confirmed the feasibility of our proposed framework, while our real-world experiment successfully identified the generalized base inertial parameter set and validated its applicability to a new robot mounting configuration.

Effective task allocation is an essential component to the coordination of heterogeneous robots. This paper proposes a hybrid task allocation algorithm that improves upon given initial solutions, for example from the popular decentralized market-based allocation algorithm, via a derivative-free optimization strategy called Speeding-Up and Slowing-Down (SUSD). Based on the initial solutions, SUSD performs a search to find an improved task assignment. Unique to our strategy is the ability to apply a gradient-like search to solve a classical integer-programming problem. The proposed strategy outperforms other state-of-the-art algorithms in terms of total task utility and can achieve near optimal solutions in simulation. Experimental results using the Robotarium are also provided.

Hovering hummingbirds have inspired small flapping-wing aerial robots. Natural flyers, including hummingbirds and bats, undergo torsional wing deformation during flapping flight owing to complex wing structure, while previous artificial wings were relatively simple and difficult to design the torsional flexibility. In this paper, we proposed a hummingbird-bat hybrid (HBH) wing in which torsional flexibility was implemented by an available fabrication technology. The HBH wing had a torsional arm at the leading edge inspired by a torsional wrist of a hummingbird. A bat-like stretchable wing membrane was also employed not to constrain the wing torsion. The membrane was supported by wing shafts of which bending stiffness was designed based on that of the feather shaft of a hummingbird. The three-dimensional (3-D) shape of the torsional arm and wing shafts was created by 3-D printing. The effect of the torsional arm and stretchable membrane on lift generation and deformation was evaluated using an electric flapping mechanism. It was confirmed that the torsional arm actually enhanced the passive wing torsion. The stretchable wing membrane further promoted the torsion effect of the torsional arm. Consequently, the HBH wing did not increase lift, but efficacy, defined as lift per input power, was greatly improved by 14% at most compared with the wing without a torsional arm.

An important component for the effective collaboration of humans with robots is the compatibility of their movements, especially when humans physically collaborate with a robot partner. Following previous findings that humans interact more seamlessly with a robot that moves with human-like or biological velocity profiles, this study examined whether humans can adapt to a robot that violates human signatures. The specific focus was on the role of extensive practice and real-time augmented feedback. Six groups of participants physically tracked a robot tracing an ellipse with profiles where velocity scaled with the curvature of the path in biological and non-biological ways, while instructed to minimize the interaction force with the robot. Three of the 6 groups received real-time visual feedback about their force error. Results showed that with 3 daily practice sessions, when given feedback about their force errors, humans could decrease their interaction forces when the robot's trajectory violated human-like velocity patterns. Conversely, when augmented feedback was not provided, there were no improvements despite this extensive practice. The biological profile showed no improvements, even with feedback, indicating that the (non-zero) force had already reached a floor level. These findings highlight the importance of biological robot trajectories and augmented feedback to guide humans to adapt to non-biological movements in physical human-robot interaction. These results have implications on various fields of robotics, such as surgical applications and collaborative robots for industry.

In this paper, we present a novel method of motion planning for performing complex manipulation tasks by using human demonstration and exploiting the screw geometry of motion. We consider complex manipulation tasks where there are constraints on the motion of the end effector of the robot. Examples of such tasks include opening a door, opening a drawer, transferring granular material from one container to another with a spoon, and loading dishes to a dishwasher. Our approach consists of two steps: First, using the fact that a motion in the task space of the robot can be approximated by using a sequence of constant screw motions, we segment a human demonstration into a sequence of constant screw motions. Second, we use the segmented screws to generate motion plans via screw-linear interpolation for other instances of the same task. The use of screw segmentation allows us to capture the invariants of the demonstrations in a coordinate-free fashion, thus allowing us to plan for different task instances from just one example. We present extensive experimental results on a variety of manipulation scenarios showing that our method can be used across a wide range of manipulation tasks.

Ensuring the safety and efficiency of human workers in environments shared with autonomous robots is of paramount importance. In this work we examine the behavior and attitudes of participants performing tasks in a noisy environment collocated with an autonomous quadcopter robot. Visual communication of spatial ownership and nonverbal (deictic gesture) requests for changes in spatial ownership are facilitated using an augmented reality (AR) head-mounted device that renders a color-keyed grid on the floor. After a request, the robot can alter floor ownership to provide participants with a safe path to complete their work. Participants (n=20n=20) in a between-subjects study took part in either a shared space condition (concurrently occupying the work floor with the robot, with obvious rationale for floor ownership) or a turn-taking condition (alternating excursions onto the grid with the robot, without apparent rationale for the floor grid colors). We find consistent evidence of potentially dangerous over-trust in the system that led to non-compliance; notably, 25% of participants intentionally walked across forbidden floor regions during the experiment. We identify design considerations and a variety of user-borne rationale for committing safety violations that designers will need to explicitly take measures to remedy in production AR safety systems.

Estimating terrain traversability in off-road environments requires reasoning about complex interaction dynamics between the robot and these terrains. However, it is challenging to create informative labels to learn a model in a supervised manner for these interactions. We propose a method that learns to predict traversability costmaps by combining exteroceptive environmental information with proprioceptive terrain interaction feedback in a self-supervised manner. Additionally, we propose a novel way of incorporating robot velocity into the costmap prediction pipeline. We validate our method in multiple short and large-scale navigation tasks on challenging off-road terrains using two different large, all-terrain robots. Our short-scale navigation results show that using our learned costmaps leads to overall smoother navigation, and provides the robot with a more fine-grained understanding of the robot-terrain interactions. Our large-scale navigation trials show that we can reduce the number of interventions by up to 57% compared to an occupancy-based navigation baseline in challenging off-road courses ranging from 400 m to 3150 m. Appendix and full experiment videos can be found in our website: https://mateoguaman.github.io/hdif.

The highly varied and deformable structure of clothing presents a challenging task in the area of robot manipulation. Recent literature has shown an increasing interest in this field, however limited information exists on the influence of end-effector selection, instead focusing on the perception, modelling, and methodology in handling fabrics. Here, we present a benchmark set of household clothing items, along with a framework for defining textile features in relation to how the objects can be grasped and manipulated. Alongside these, we present four example benchmarks for evaluating the performance of a robot end-effector in relation to the grasping and manipulation of common pieces of clothing: Edge drag accuracy, edge grasp resilience, grasp encapsulation, and grasp fold generation. We perform these benchmarks on several common robot end-effectors (Franka Emika (FE) Hand with standard and Fin Ray® style fingers (Flex), Robotiq 2F-140, and the Openhand Model T42) and present and discuss their respective performances. Results show that the Robotiq scored highest across most benchmarks, closely followed by the FE hand. The T42 showed excellent encapsulation of items, while the FE (Flex) was particularly successful picking up flat edges.