IROS 2021 Papers

This paper presents the design, fabrication, and preliminary results of a soft hip exosuit to assist hip flexion and extension during walking. The exosuit uses soft and compliant materials to create a wearable robot that has a low profile, low mass, and is highly flexible to freely move with the user’s natural range of motion. The Soft Robotic Hip Exosuit (SR-HExo) consists of flat fabric pneumatic artificial muscles (ff-PAM) that contract when pressurized. The ff-PAM actuators are oriented in an ‘X’ shape to allow for natural range of motion across the hip joint and can generate 190 N of force at 200 kPa in a 0.3 sec window. The ‘X’ configuration (X-ff-PAM) actuators were placed on the anterior and posterior sides of the hip joint with height adjustable Velcro straps. Extension assistance and flexion assistance was provided in 10-45% and 50-90% of the gait cycle, respectively. To evaluate the effectiveness of the SR-HExo with healthy participants, hip range of motion and muscle activity during walking were monitored using a motion capture system and surface electromyography sensors. The impact of the SR-HExo on the range of motion was minimal with only a 4.0o reduction from the target range of motion of 30o. The exosuit contributed to reducing hip muscle activity. Hip extensor muscles showed a reduction of 13.1% for the gluteus maximus and 6.6% for the biceps femoris. Hip flexor muscles showed a reduction of 10.7% for the iliacus and 27.7% for the rectus femoris.

The use of assistive technologies in industrial environments to improve human ergonomics and comfort in repetitive and high effort tasks have increased considerably in the last decade. Predominantly, the goal is to provide additional physical support through lightweight and wearable devices, without posing major constraints to the human body movements. Towards achieving this objective, in this work we present a novel actuation mechanism for a soft assistive device, by taking into account the human elbow torque-angle profile. The proposed design integrates a single motor coupled with an elastic bungee and a cam-spool mechanism to enable energy exchange during the elbow flexion movement, while allowing for free-motions during the extension of the joint. A cable-driven transmission with passive elastic attachments is employed to implement compliant couplings with the wearer and to achieve easy donning/doffing. Experiments are conducted on two 3D printed functional prototypes. Results suggest that the assistive elbow torque is effectively transmitted with an average 90% success for balancing a 5N payload, and the free-motion range of 108° is measured for both flexion and extension.

Detecting dynamic objects and predicting static road information such as drivable areas and ground heights are crucial for safe autonomous driving. Previous works studied each perception task separately, and lacked a collective quantitative analysis. In this work, we show that it is possible to perform all perception tasks via a simple and efficient multi-task network. Our proposed network, LidarMTL, takes raw LiDAR point cloud as inputs, and predicts six perception outputs for 3D object detection and road understanding. The network is based on an encoder-decoder architecture with 3D sparse convolution and deconvolution operations. Extensive experiments verify the proposed method with competitive accuracies compared to state-of-the-art object detectors and other task-specific networks. LidarMTL is also leveraged for online localization. Code and pre-trained model have been made available at https://github.com/frankfengdi/LidarMTL.

Differential mechanisms allow the designers of robotic and prosthetic grippers and hands to create devices that require a minimal number of motors in order to grasp a plethora of everyday life objects, leading to light-weight, compact, and low-cost implementations. The working principle of differential mechanisms is simple. They allow the distribution of the forces exerted by a single actuator to multiple outputs (e.g., fingers). This reduction in the number of motors leads to underactuation, which is the use of fewer motors than the available degrees of freedom. But differentials need also to be power-efficient, compact, adaptive, and lightweight. Most of the existing solutions lack at least one of these attributes. In this paper, we focus on the design, modeling, and development of a compact, adaptive, series elastic differential. The proposed mechanism consists of four elastic elements connected in series with the four output attachments. The compression of the elastic elements during grasping allows the gripper or hand to conform to the object’s shape. The efficiency of the differential mechanism is experimentally validated using two different types of experiments, measuring: i) the maximum achievable tension load at the outputs, and ii) the maximum achievable compliance of a single output when all other outputs are blocked. The proposed differential has been employed for the development of a gripper and its efficiency has been assessed by executing grasping tasks with several everyday life objects. The device can be easily replicated using additive manufacturing and off-the-shelf materials and is disseminated in an open-source manner.

Shape Memory Alloys (SMAs) are a type of smart material that are commonly used in compact and lightweight linear actuators. Paired with compliant motion conversion mechanisms, these alloys can be transformed into lightweight grippers ideal for applications such as drone deliveries. This work proposes to use the inherent stiffness of a compliant mechanism with an SMA coil, to design and size a self-biasing SMA gripper. In this paper, with the help of 2.5D design, a functional radial 4-prong gripper is designed and fabricated. While only weighing 17g, it has a gripping force and stroke of 1.8N and 4mm, respectively. Along with this functional prototype, an analytical model and design framework is presented allowing for the design and sizing of similar self-biasing SMA actuators for other different light-weight applications.

In this article we propose a distributed collision avoidance scheme for multi-agent unmanned aerial vehicles (UAVs) based on nonlinear model predictive control (NMPC), where other agents in the system are considered as dynamic obstacles with respect to the ego agent. Our control scheme operates at a low level and commands roll, pitch and thrust signals at a high frequency, each agent broadcasts its predicted trajectory to the other ones, and we propose an obstacle prioritization scheme based on the shared trajectories to allow up-scaling of the system. The NMPC problem is solved using an embedded solver generated by Optimization Engine (OpEn) where PANOC is combined with an augmented Lagrangian method to compute collision-free trajectories. We evaluate the proposed scheme in several challenging laboratory experiments for up to ten aerial agents, in dense aerial swarms.

We present a framework for planning complex motor actions such as pouring or scooping from arbitrary start states in cluttered real-world scenes. Traditional approaches to such tasks use dynamic motion primitives (DMPs) learned from human demonstrations. We enhance a recently proposed state-of-the-art DMP technique capable of obstacle avoidance by including them within a novel hybrid framework. This complements DMPs with sampling-based motion planning algorithms, using the latter to explore the scene and reach promising regions from which a DMP can successfully complete the task. Experiments indicate that even obstacle-aware DMPs suffer in task success when used in scenarios which largely differ from the trained demonstration in terms of the start, goal, and obstacles. Our hybrid approach significantly outperforms obstacle-aware DMPs by successfully completing tasks in cluttered scenes for a pouring task in simulation. We further demonstrate our method on a real robot for pouring and scooping tasks.

In collaborative robotic applications, human and robot have to work together to accomplish a common job, composed by a set of tasks. In order to achieve an efficient human-robot collaboration (HRC), it is important to have an integration between a proper task scheduling strategy and a task execution strategy. The first must deal with the variability of the two agents, while the second must deal with the safety standards. In this paper, we propose an integrated architecture for task scheduling and execution in a collaborative cell. The tasks are dynamically scheduled handling the uncertainity in both the human and the robot behaviors. Subsequently, at the execution level, the task is accomplished computing trajectories comply with the safety regulations. The planning information are mutually integrated in real-time with the scheduling procedure in order improve the HRC.

Hydrostatic transmission has shown promising results for enabling the manipulator to achieve low effective inertia, high stiffness, and high torque density. However, the incompressibility of fluid causes the lack of compliance, so that it could not provide intrinsic safety. Thus, it would be advantageous to introduce series compliance on the hydrostatic manipulator for adjusting stiffness depending on the situation. Here, we developed a safe and high-performance hydrostatic actuator based on the switchable stiffness mechanism implemented with an air spring and solenoid valve. The hydrostatic transmission is implemented with rolling diaphragms to attain zero fluid leakage and low seal friction. Air spring is serially connected to hydraulic lines for achieving compliance. Its modes (i.e., stiff and compliant modes) can be rapidly switched by modulating water flow via the solenoid valve. Block stiffness experiment shows that the stiffness of stiff mode is 9.63 times stiffer than one of compliant mode at 100 kPa. We experimentally demonstrated that compliant mode could mitigate the impact force to the level that a tangerine would not be crushed. The stiffness was switched within 12 ms; hence it is fast enough to be used for feedback application with vision or tactile sensors. As a result, the developed actuator can ensure safety without sacrificing the dynamic performance, owing to the simple and rapidly switchable stiffness mechanism.

Object detection and semantic segmentation are two of the most widely adopted deep learning algorithms in agricultural applications. One of the major sources of variability in image quality acquired outdoors for such tasks is changing lighting conditions that can alter the appearance of the objects or the contents of the entire image. While transfer learning and data augmentation reduce the need for large amount of data to train deep neural networks to some extent, the large variety of cultivars and the lack of shared datasets in agriculture makes wide-scale field deployments difficult. In this paper, we present an active lighting-based camera system that generates robust and uniform images in any lighting conditions. We provide extensive validation experiments to evaluate the consistency in the quality of the images. Metrics for assessing image uniformity like Structural Similarity (SSIM) index and the Peak Signal to Noise Ratio (PSNR) ranged from 83.78 to 93.79 and from 25.30 to 31.78 respectively, showing stability over a day with changing sunlight. The validation stage also showed that the generated images effectively reduce the amount of training samples for object detection using deep neural networks. The camera system was then deployed in real field experiments for counting buds in dormant vines and shoots in early season grape vines as well as for counting apples in orchards. The mean absolute errors obtained when compared to ground truth were 5%, 2.55% and 8.57%, respectively.

In this paper, we propose a data-driven modelling framework using a sparse regression technique to find the governing equations of dynamics systems. With this approach, the prior knowledge of features from simple structures can be used to deduce which on complex structures. The prior knowledge of single-pendulums, double-pendulums, and spherical pendulum enlightens the guess of the feature library for a 3-DOF manipulator. The feature library is sparsifled with a fully autonomous machine learning algorithm composited of the L1-regularization and proportional filter. The training dataset with non-zero-mean Gaussian noise simulates real-world conditions and proves the system's robustness to the noise. Compared with the neural-network-based system identification method, this paper's technique can be promptly applied in dynamic trajectory planning. A simulation of the optimal trajectory planning for the obstacle-avoidance on the Lynxmotion robot is accomplished by optimizing the objective function constructed with energy and penalty function. Results of the simulation support that the estimated model works correctly. Comparison between the data-driven and the closed-form model evidences the reliability and robustness of this identification technique.

We study a stochastic version of the classic orienteering problem where the time to traverse an edge is a continuous random variable. For a given temporal deadline B, our solution produces a policy, i.e., a function that, based on the current position along a solution path and the elapsed time, decides whether to continue along the path or take a shortcut to avoid missing the deadline. The solution is based on a formulation using constrained Markov decision processes to ensure that the deadline is met with a preassigned confidence level. To expedite the computation, a Monte Carlo simulation on an open loop policy is run to determine how to adaptively discretize the temporal dimension and therefore reduce the number of states and the number of optimization variables in the associated linear program. Our results show that the adaptive algorithm matches the performance of the non-adaptive one while taking significantly less time.

In the field of large-scale SLAM for autonomous driving and mobile robotics, 3D point cloud based place recognition has aroused significant research interest due to its robustness to changing environments with drastic daytime and weather variance. However, it is time-consuming and effort-costly to obtain high-quality point cloud data for place recognition model training and ground truth for registration in the real world. To this end, a novel registration-aided 3D domain adaptation network for point cloud based place recognition is proposed. A structure-aware registration network is introduced to help to learn features with geometric information and a 6-DoFs pose between two point clouds with partial overlap can be estimated. The model is trained through a synthetic virtual LiDAR dataset through GTA-V with diverse weather and daytime conditions and domain adaptation is implemented to the real-world domain by aligning the global features. Our results outperform state-of-the-art 3D place recognition baselines or achieve comparable on the real-world Oxford RobotCar dataset with the visualization of registration on the virtual dataset.

Prolonged remote tele-locomanipulation of multi degrees-of-freedom mobile manipulators requires a compromise between the system’s performance and the operator’s ergonomics. Neglecting this demand can significantly affect either the task completion or the level of comfort to achieve it. However, the simultaneous consideration of these key factors has received less attention in the literature. To respond to this demand, in this work, we introduce a new teleoperation setup, which integrates the features of an ergonomic and a highly maneuverable interface into a unified solution. The ergonomic part of the interface implements a 3D mouse-like functionality, enabling the execution of long navigation tasks for the floating base. The highly manoeuvrable interface instead, enables the operator to perform dynamic or more precise manipulation by moving his/her arm in space. The locomotion and manipulation modes of the follower robot are controlled separately, which can be easily and seamlessly switched by the operator by pressing a button at any moment. Furthermore, due to the follower manipulator’s redundancy, this robot is controlled by a hierarchical quadratic programming technique which enables the definition of a set of secondary tasks to be executed in the robot’s nullspace. Finally, to demonstrate the advantages and disadvantages of the proposed user interfaces, five participants are asked to perform two different experiments: (i) target selection task on a moving surface and (ii) remote path tracking on a fixed surface. The quantitative and qualitative analyses show the effectiveness of the proposed interface during the teleoperation tasks, especially when it comes to the precise and dynamic task execution.

In this paper, we propose an automatic robot-rat interaction framework that enables a robotic rat to realize real-time localization, tracking and movement analysis of a laboratory rat. Specifically, we combine an object detector with stereo matching to achieve fast localization of the laboratory rat. Combined with the rat-like motion of the robot, one-step tracking of the rat is achieved, which enables the robot to eliminate last location error in one cycle of visual servo control. When positioning the rat, a unified quantitative description and analysis of rat motion is implemented by using a state vector composed of the centroids of head, body and tail. Preliminary robot-rat interaction tests show that the robot achieved a steady tracking of a fast-moving rat for a duration of 10 minutes. To the best of our knowledge, it is the first time that a rat-sized robot achieves a continuous tracking of actual rats by a built-in miniature stereo vision system. Experimental results show that the sequence of state vectors accurately represents the pitch movement of the rat. Thus, this work is a step toward more natural interaction between robots and animals.

In the process of landing unmanned aerial vehicles (UAVs) on an unmanned surface vehicle (USV), a manipulator can be applied to help the UAV land safely and accurately. However, it is a challenge to control the manipulator on a disturbed USV due to joint velocity constraints and bandwidth limitations. To solve this problem, a predictive control framework is proposed in this paper. We leverage a first-order delay system to describe the kinematics of each joint, and control joint velocities by the model predictive controller (MPC). To generate references for MPC, the motion of the floating base needs to be predicted. We apply the recent approach for motion prediction based on the wavelet network (WN) and modify the network to get smooth trajectories. The accuracy of the modified wavelet network (MWN) for motion prediction is tested on four-hour motion data from the real ocean environment and the smoothness of the generated trajectories is also evaluated. Simulations and experiments are implemented to verify the proposed method, the results show that the average control accuracies are improved by more than 30% and 50% in position and rotation compared with the traditional inverse kinematics (IK) controller for 1 Hz base fluctuation.

Ankle joints play key roles in everyday locomotion, such as walking, stair climbing, and sit-to-stand. Despite the achievement in designing powered prosthetic ankles, engineers still face challenges to duplicate the full mechanics of ankle joints, including high torque, large range of motion (ROM), low profile, backdrivability, and efficiency, using electric motors and related transmissions. In this study, our goal was to develop a new active prosthetic ankle, Variable Spring embedded Motor-ball screw (VSeM) ankle, to meet all these requirements at the same time. Using a manually adjustable elastic element, which is parallel with our motor actuator, we can readjust the ROM of VSeM to handle all normal locomotion tasks. VSeM’s capability to mimic human ankle was validated through both bench tests and human subject tests.

Non-contact manipulation technology has extensive application in the manipulation and fabrication of micro/nanomaterials. However, the manipulation devices are often precise and complex, operated only by professionals and subject to site constraints. We propose a simple optoelectronic tweezer platform, which can be controlled remotely and simply for the manipulation of microparticles at different scales, based on the novel manipulation technique called optically-induced dielectrophoresis. In this work, we design and set up the optoelectronic tweezer manipulation platform and develop the full-function human-computer interactive control interface and graphics rendering system to simplify the micro-operation process. Using Qt5.0 development environment, an experimental image processing system with multi-thread characteristics is developed, and the information interaction requirements needed in the experimental operation of optoelectronic tweezers are integrated into a control system to achieve unified information management and data analysis. Combined with cloud computing technology, the system realizes local/remote synchronous linkage operation, with cross-platform operation capability of various portable operating terminals like laptop and iPad.

Suitable datasets are an integral part of robotics research, especially for training neural networks in robot perception. However, in many domains, suitable real-world data are scarce and cannot be easily obtained. This problem is especially prevalent for unstructured outdoor environments, in particular, planetary ones. Recent advances in photorealistic simulations help researchers to simulate close-to-real data in many domains. Yet, there exists no high-quality synthetic data for planetary exploration tasks. Also, existing simulators lack the fidelity required for generating planetary data, which is inherently less structured than human environments. Synthetic planetary data requires careful modeling and annotation of many different terrain aspect and details, such as textures and distributions of rocks, to become a valuable test-bed for robotics. To fill this gap, we present a novel simulator specifically designed for the needs of planetary robotics visual tasks, but also applicable for other outdoor environments. Our simulator is capable of generating large varieties of (planetary) outdoor scenes with rich generation of meta data, such as multilevel semantic and instance annotations. To demonstrate the wide applicability of this new simulator, we evaluate its performance on typical robotics applications, i.e. semantic segmentation, instance segmentation, and visual SLAM. Our simulator is accessible under https://github.com/DLR-RM/oaisys.