ICRA 2023 论文列表

In dynamic environments, robotic manipulators and especially cobots must be able to react to changing circumstances while in motion. This substantiates the need for quick trajectory planning algorithms that are able to cope with arbitrary velocity and acceleration boundary conditions. Apart from dynamic re-planning, being able to seamlessly join trajectories together opens the door for divide-and-conquer-type algorithms to focus on the individual parts of a motion separately. While geodesic motion planning has proven that it can produce very smooth and efficient actuator movement, the problem of incorporating non-zero boundary conditions has not been addressed yet. We show how a set of generalized coordinates can be used to transition between boundary conditions and free movement in an optimal way while still retaining the known advantages of geodesic planners. We also outline, how our approach can be combined with the family of time-scaling algorithms for further improvement of the generated trajectories.

In this paper, we propose a novel dynamic gait controller for the repetitive behavior of soft robot manipulators performing routine tasks. Compliance with soft robots is advantageous when the robot interacts with living organisms and other fragile objects. However, predicting and controlling repetitive behavior is challenging because of hysteresis and non-linear dynamics governing the interactions. Existing priorfree methods track the dynamic state using recurrent neural networks or rely on known generalized coordinates describing the robot's state. We propose to model the interaction induced by the repetitive behavior as gait dynamics and represent the dynamic state with Central Pattern Generator (CPG) tracking the motion phase and thus reduce the complexity of the robot's forward model. The proposed method bootstraps an ensemble of the forward models exploring multiple dynamic contexts that are expanded as it searches for repetitive motion producing the target repetitive behavior. The proposed approach is experimentally validated on a pneumatically actuated soft robot arm I-Support, where the method infers gaits for different targets.

One recent promising approach to the Visual Place Recognition (VPR) problem has been to fuse the place recognition estimates of multiple complementary VPR techniques using methods such as shared representative appearance learning (SRAL) and multi-process fusion. These approaches come with a substantial practical limitation: they require all potential VPR methods to be brute-force run before they are selectively fused. The obvious solution to this limitation is to predict the viable subset of methods ahead of time, but this is challenging because it requires a predictive signal within the imagery itself that is indicative of high performance methods. Here we propose an alternative approach that instead starts with a known single base VPR technique, and learns to predict the most complementary additional VPR technique to fuse with it, that results in the largest improvement in performance. The key innovation here is to use a dimensionally reduced difference vector between the query image and the top-retrieved reference image using this baseline technique as the predictive signal of the most complementary additional technique, both during training and inference. We demonstrate that our approach can train a single network to select performant, complementary technique pairs across datasets which span multiple modes of transportation (train, car, walking) as well as to generalise to unseen datasets, outperforming multiple baseline strategies for manually selecting the best technique pairs based on the same training data.

The recent multi-modality models have achieved great performance in many vision tasks because the extracted features contain the multi-modality knowledge. However, most of the current registration descriptors have only concentrated on local geometric structures. This paper proposes a method to boost point cloud registration accuracy by transferring the multi-modality knowledge of pre-trained multi-modality model to a new descriptor neural network. Different to the previous multi-modality methods that requires both modalities, the proposed method only requires point clouds during inference. Specifically, we propose an ensemble descriptor neural network combining pre-trained sparse convolution branch and a new point-based convolution branch. By fine-tuning on a single modality data, the proposed method achieves new state-of-the-art results on 3DMatch and competitive accuracy on 3DLoMatch and KITTI. The code and the trained model will be released at https://github.com/phdymz/DBENet.git.

The use of Micro Aerial Vehicles (MAVs) for inspection and surveillance missions has proved to be extremely useful, however, their usability is negatively impacted by the large power requirements and the limited operating time. This work describes the design and development of a novel hybrid aerial-ground vehicle, enabling multi-modal mobility and long operating time, suitable for long-endurance inspection and monitoring applications. The design consists of a MAV with two tiltable axles and four independent passive wheels, allowing it to fly, approach, land and move on flat and inclined surfaces, while using the same set of actuators for all modes of locomotion. In comparison to existing multi-modal designs with passive wheels, the proposed design enables a higher ground locomotion efficiency, provides a higher payload capacity, and presents one of the lowest mass increases due to the ground actuation mechanism. The vehicle's performance is evaluated through a series of real experiments, demonstrating its flying, ground locomotion and wall-climbing capabilities, and the energy consumption for all modes of locomotion is evaluated.

For humanoids to be deployed in demanding situations, such as search and rescue, highly intelligent decision making and proficient sensorimotor skill is expected. A promising solution is to leverage human prowess by interconnecting robot and human via teleoperation. Towards creating seamless operation, this paper presents a dynamic telelocomotion framework that synchronizes the gait of a human pilot with the walking of a bipedal robot. First, we introduce a method to generate a virtual human walking model from the stepping behavior of a human pilot which serves as a reference for the robot to walk. Second, the dynamics of the walking reference and robot walking are synchronized by applying forces to the human pilot and the robot to achieve dynamic similarity between the two systems. This enables the human pilot to continuously perceive and cancel any asynchrony between the walking reference and robot. A consistent step placement strategy for the robot is derived to maintain dynamic similarity through step transitions. Using our human-machine-interface, we demonstrate that the human pilot can achieve stable and synchronous teleoperation of a simulated robot through stepping-in-place, walking, and disturbance rejection experiments. This work provides a fundamental step towards transferring human intelligence and reflexes to humanoid robots.

We present SunBot, a robotic system for the study and implementation of fish-inspired tearing manipulations. Various fish species–such as the sunburst butterflyfish-feed on prey fixed to substrates, a maneuver previously not demonstrated by robotic fish which typically specialize for open water swimming and surveillance. Biological studies indicate that a dynamic “head flick” behavior may play a role in tearing off soft prey during such feeding. In this work, we study whether the robotic tail is an effective means to generate such head motions for ungrounded tearing manipulations in water. We describe the function of SunBot and compare the forces that it applies to a fixed prey in the lab while varying tail speeds and ranges of motion. A simplified dynamic template model for the tail-driven head flick maneuver matches peak force magnitudes from experiments, indicating that inertial effects of the fish's body play a substantial role. Finally, we demonstrate a tearing scenario and evaluate a free-swimming trial of SunBot – this is important to show that the actuator that enables swimming also provides the new dual purpose of forceful tearing manipulation.

Forest canopies are vital ecosystems, but remain understudied due to difficult access. Forests could be monitored with a network of biodegradable sensors that break down into environmentally friendly substances at the end of their life. As a first step in this direction, this paper details the development of a biodegradable origami gripper to attach conventional sensors to branches, deployable with an aerial robot. Through exposure to sufficient moisture the gripper loses contractile force, dropping the sensor to the ground for easier collection. The origami design of the gripper as well as biodegradable materials selection is detailed, allowing for further extensions utilizing biodegradable origami. Both the gripper and the gelatin hydrogel used as an actuating elastic element for generating the grasping force are experimentally characterized, with the gripper demonstrating a maximum holding force of 1 N. Additionally, the degradation of the gripper until failure in the presence of moisture is also investigated, where the gripper can absorb up to 10 ml of water before falling off a branch. Finally, deployment of the gripper on a tree branch with an aerial robot is demonstrated. Overall, the biodegradable origami gripper represents a first step towards a more scalable and environmentally sustainable approach for ecosystem monitoring.

Learning from demonstration (LfD) with behavior cloning is attractive for its simplicity; however, compounding errors in long and complex skills can be a hindrance. Considering a target skill as a sequence of motor primitives is helpful in this respect. Then the requirement that a motor primitive ends in a state that allows the successful execution of the subsequent primitive must be met. In this study, we focus on this problem by proposing to learn an explicit correction policy when the expected transition state between primitives is not achieved. The correction policy is learned via behavior cloning by the use of Conditional Neural Motor Primitives (CNMPs) that can generate correction trajectories in a context-dependent way. The advantage of the proposed system over learning the complete task as a single action is shown with a table-top setup in simulation, where an object has to be pushed through a corridor in two steps. Then, the applicability of the proposed method to bi-manual knotting in the real world is shown by equipping an upper-body humanoid robot with the skill of making knots over a bar in 3D space.

Wearable robotic exoskeletons hold great promise for gait rehabilitation as portable, accessible tools. However, a better understanding of the potential for exoskeletons to elicit neural adaptation-a critical component of neurological gait rehabilitation-is needed. In this study, we investigated whether humans adapt to bilateral asymmetric stiffness perturbations applied by a hip exoskeleton, taking inspiration from the asymmetry augmentation strategies used in split-belt treadmill training. During walking, we applied torques about the hip joints to repel the thigh away from a neutral position on the left side and attract the thigh toward a neutral position on the right side. Six participants performed an adaptation walking trial on a treadmill while wearing the exoskeleton. The exoskeleton elicited time-varying changes and aftereffects in step length and propulsive/braking ground reaction forces, indicating behavioral signatures of neural adaptation. These responses resemble typical responses to split-belt treadmill training, suggesting that the proposed intervention with a robotic hip exoskeleton may be an effective approach to (re)training symmetric gait.

This paper presents a novel two-step method for synthesizing spatial linkage mechanisms. Compared with planar mechanisms, the main challenge in synthesizing spatial mechanisms is that the generating motion varies depending on its mechanism topologies. Therefore, we propose a big data approach to determine the topology of spatial mechanisms. We adopt a three-dimensional (3D) spring-connected rigid block model to represent the topology of the spatial mechanism and project 3D motion onto three orthogonal planes to determine the mechanism topology with big data. In addition, a gradient-based dimension synthesis procedure was carried out to determine a detailed dimension using already determined mechanism topology by mechanism big data. Also, several successful case studies by the proposed approach are presented to support the effectiveness of the proposed synthesis method.

In medical robotics and image-guided surgery (IGS), registration is needed in order to align together the coordinate frames of robots, medical imaging modalities, surgical tools, and patients. Existing registration algorithms often assume one point set to be a noise-free model while the other to contain noise and outliers. However, in real scenarios, noise and outliers can exist in both point sets to be registered. To eliminate the above-mentioned challenge, in this paper, we formally formulate the Bi-directional Generalised Rigid Point Set Registration (Bi-GRPSR) problem where normal vectors are adopted, bi-directional probability density function (PDFs) and Hybrid Mixture Models (HMMs) are constructed to derive the objective function. Bi-GRPSR considering anisotropic positional noise is thus cast as a maximum likelihood estimation (MLE) problem, which is solved by the proposed Bi-directional Generalised Anisotropic Coherent Point Drift (Bi-AGCPD) where spatially nearby points are considered to move coherently and iterative expectation maximization (EM) steps are involved. Experimental results on two human bone point sets, under different settings of noise, outliers, and overlapping ratios, validate the effectiveness and improvements of Bi-AGCPD over existing probabilistic and learning-based methods.

In this paper, we propose a strategy for robot-assisted dressing with multi-component garments, such as gloves. Most studies in robot-assisted dressing usually experiment with single-component garments, such as sleeves, while multi-component tasks are often approached as sequential single-component problems. In dressing scenarios with more complex garments, robots should estimate the alignment of the human body to the manipulated garments, and revise their dressing strategy. In this paper, we focus on a glove dressing scenario and propose a decision process for selecting dressing action primitives on the different components of the garment, based on a hierarchical representation of the task and a set of environmental conditions. To complement this process, we propose a set of bi-manual control strategies, based on hybrid position, visual, and force feedback, in order to execute the dressing action primitives with the deformable object. The experimental results validate our method, enabling the Baxter robot to dress a mannequin's hand with a gardening glove.

Deep reinforcement learning (RL) has brought many successes for autonomous robot navigation. However, there still exists important limitations that prevent real-world use of RL-based navigation systems. For example, most learning approaches lack safety guarantees; and learned navigation systems may not generalize well to unseen environments. Despite a variety of recent learning techniques to tackle these challenges in general, a lack of an open-source benchmark and reproducible learning methods specifically for autonomous navigation makes it difficult for roboticists to choose what learning methods to use for their mobile robots and for learning researchers to identify current shortcomings of general learning methods for autonomous navigation. In this paper, we identify four major desiderata of applying deep RL approaches for autonomous navigation: (D1) reasoning under uncertainty, (D2) safety, (D3) learning from limited trial-and-error data, and (D4) generalization to diverse and novel environments. Then, we explore four major classes of learning techniques with the purpose of achieving one or more of the four desiderata: memory-based neural network architectures (D1), safe RL (D2), model-based RL (D2, D3), and domain randomization (D4). By deploying these learning techniques in a new open-source large-scale navigation benchmark and real-world environments, we perform a comprehensive study aimed at establishing to what extent can these techniques achieve these desiderata for RL-based navigation systems.

The main challenge in developing effective reinforcement learning (RL) pipelines is often the design and tuning the reward functions. Well-designed shaping reward can lead to significantly faster learning. Naively formulated rewards, however, can conflict with the desired behavior and result in overfitting or even erratic performance if not properly tuned. In theory, the broad class of potential based reward shaping (PBRS) can help guide the learning process without affecting the optimal policy. Although several studies have explored the use of potential based reward shaping to accelerate learning convergence, most have been limited to grid-worlds and low-dimensional systems, and RL in robotics has predominantly relied on standard forms of reward shaping. In this paper, we benchmark standard forms of shaping with PBRS for a humanoid robot. We find that in this high-dimensional system, PBRS has only marginal benefits in convergence speed. However, the PBRS reward terms are significantly more robust to scaling than typical reward shaping approaches, and thus easier to tune.

Degraded visual environments have strong impacts on the quality of LiDAR data. Experiments in artificial fog conditions show that noise points caused by water particles present various distance distributions which depend on visibility. This article introduces a mathematical framework based on Bayesian inference and Markov Chain Monte-Carlo sampling to infer optical visibility from point clouds. The visibility estimation is cast as a classification problem based on the identification of the distance distributions. Contrary to deep learning methods, our approach is model-based and focuses on the design of a full probabilistic framework, more comprehensible, which is critical for autonomous driving. Ultimately, the impact of the optical visibility on the probability of detection of standard targets is assessed, which can yield improvements on autonomous vehicles performances in adverse weather conditions.

Affordances are a fundamental concept in robotics since they relate available actions for an agent depending on its sensory-motor capabilities and the environment. We present a novel Bayesian deep network to detect affordances in images, at the same time that we quantify the distribution of the aleatoric and epistemic variance at the spatial level. We adapt the Mask-RCNN architecture to learn a probabilistic representation using Monte Carlo dropout. Our results outperform the state-of-the-art of deterministic networks. We attribute this improvement to a better probabilistic feature space representation on the encoder and the Bayesian variability induced at the mask generation, which adapts better to the object contours. We also introduce the new Probability-based Mask Quality measure that reveals the semantic and spatial differences on a probabilistic instance segmentation model. We modify the existing Probabilistic Detection Quality metric by comparing the binary masks rather than the predicted bounding boxes, achieving a finer-grained evaluation of the probabilistic segmentation. We find aleatoric variance in the contours of the objects due to the camera noise, while epistemic variance appears in visual challenging pixels.

Patrolling with multiple robots is a challenging task. While the robots collaboratively and repeatedly cover the regions of interest in the environment, their routes should satisfy two often conflicting properties: i) (efficiency) the time intervals between two consecutive visits to the regions are small; ii) (unpredictability) the patrolling trajectories are random and unpredictable. We manage to strike a balance between the two goals by i) recasting the original patrolling problem as a Graph Deep Learning problem; ii) directly solving this problem on the graph in the framework of cooperative multi-agent reinforcement learning. Treating the decisions of a team of agents as a sequence input, our model outputs the agents' actions in order by an autoregressive mechanism. Extensive simulation studies show that our approach has comparable performance with existing algorithms in terms of efficiency and outperforms them in terms of unpredictability. To our knowledge, this is the first work that successfully solves the patrolling problem with reinforcement learning on a graph.

Typical algorithms for point cloud registration such as Iterative Closest Point (ICP) require a favorable initial transform estimate between two point clouds in order to perform a successful registration. State-of-the-art methods for choosing this starting condition rely on stochastic sampling or global optimization techniques such as branch and bound. In this work, we present a new method based on Bayesian optimization for finding the critical initial ICP transform. We provide three different configurations for our method which highlights the versatility of the algorithm to both find rapid results and refine them in situations where more runtime is available such as offline map building. Experiments are run on popular data sets and we show that our approach outperforms state-of-the-art methods when given similar computation time. Furthermore, it is compatible with other improvements to ICP, as it focuses solely on the selection of an initial transform, a starting point for all ICP-based methods.