ICRA 2020 论文列表

Spiking Neural Networks (SNNs) are bio-inspired networks that process information conveyed as temporal spikes rather than numeric values. An example of a sensor providing such data is the event-camera. It only produces an event when a pixel reports a significant brightness change. Similarly, the spiking neuron of an SNN only produces a spike whenever a significant number of spikes occur within a short period of time. Due to their spike-based computational model, SNNs can process output from event-based, asynchronous sensors without any pre-processing at extremely lower power unlike standard artificial neural networks. This is possible due to specialized neuromorphic hardware that implements the highly-parallelizable concept of SNNs in silicon. Yet, SNNs have not enjoyed the same rise of popularity as artificial neural networks. This not only stems from the fact that their input format is rather unconventional but also due to the challenges in training spiking networks. Despite their temporal nature and recent algorithmic advances, they have been mostly evaluated on classification problems. We propose, for the first time, a temporal regression problem of numerical values given events from an event-camera. We specifically investigate the prediction of the 3- DOF angular velocity of a rotating event-camera with an SNN. The difficulty of this problem arises from the prediction of angular velocities continuously in time directly from irregular, asynchronous event-based input. Directly utilising the output of event-cameras without any pre-processing ensures that we inherit all the benefits that they provide over conventional cameras. That is high-temporal resolution, high-dynamic range and no motion blur. To assess the performance of SNNs on this task, we introduce a synthetic event-camera dataset generated from real-world panoramic images and show that we can successfully train an SNN to perform angular velocity regression.

Search and rescue robotics is becoming a relevant topic in the last years and the growing number of robotic platforms and dedicated projects is the evidence of the interest in this area. In this context, the possibility to drive a remote robot with an exoskeleton is a promising strategy to enhance dexterity, reduce operator effort and save time. However, the use of haptic feedback (bilateral teleoperation) may lead to instability in the presence of communication delay and more complex is the case of bimanual teleoperation where the two arms can exchange energy. In this work, we present a bimanual teleoperation system based on an exoskeletal master, where multi-degrees of freedom (multi-DoFs) and kinematically different devices are involved. In the implemented architecture the two slaves are managed in parallel and independently passivated using the Time Domain Passivity Approach (TDPA) extended for multi-DoFs devices. To investigate the stability of the architecture we designed two tasks highly related to real disaster scenarios: the first one was useful to verify the system behavior in case of small movements and constrained configurations, whereas the second experiment was designed to involve larger contact forces and movements. Moreover, we compared the effect of both delay and low control loop frequency on the stability of the system when TDPA was applied. From the results, it was evident that the overall system exhibited a stable behavior with the use of the TDPA, even passivating the two slaves independently, under simulated time delay and in presence of a low control loop frequency.

In conventional robotic surgery, the manipulating methods exhibit limitations that are strongly related to the advantages and disadvantages of a pinch grip and power grip. The context of this study is focused on the introduction of a combined grip to compensate for such restraints. In particular, this study proposed the combined-grip-handle scheme on a master manipulator. In this framework, the position of fingertips was designed to be adjustable in distance and direction to allow for a pinch grip motion around the holding axis of a power grip motion. A ring-bar experiment applying the master-slave scheme was conducted with the master manipulator under several manipulating conditions of the combined grip and the conventional gripping types. Results for using the combined grip demonstrated that the combined grip showed better performance on the positioning operation, compared with the conventional gripping types.

The latency in vision-based sensor systems used in human-robot collaborative environments is an important safety parameter which in most cases has been neglected by researchers. The main reason for this neglect is the lack of an accurate ground-truth sensor system with a minimal delay to benchmark the vision-sensors against. In this paper the latencies of 3D vision-based sensors are experimentally evaluated and analyzed using an accurate laser-tracker system which communicates on a dedicated EtherCAT channel with minimal delay. The experimental results in the paper demonstrate that the latency in the vision-based sensor system is many orders higher than the latency in the control and actuation system.

Individuals living with paralysis or amputation can operate robotic prostheses using input signals based on their intent or attempt to move. Because sensory function is lost or diminished in these individuals, haptic feedback must be non-collocated. The intracortical brain computer interface (iBCI) has enabled a variety of neural prostheses for people with paralysis. An important attribute of the iBCI is that its input signal contains signal-independent noise. To understand the effects of signal-independent noise on a system with non-collocated haptic feedback and inform iBCI-based prostheses control strategies, we conducted an experiment with a conventional haptic interface as a proxy for the iBCI. Ablebodied users were tasked with locating an indentation within a virtual environment using input from their right hand. Non-collocated haptic feedback of the interaction forces in the virtual environment was augmented with noise of three different magnitudes and simultaneously rendered on users' left hands. We found increases in distance error of the guess of the indentation location, mean time per trial, mean peak absolute displacement and speed of tool movements during localization for the highest noise level compared to the other two levels. The findings suggest that users have a threshold of disturbance rejection and that they attempt to increase their signal-to-noise ratio through their exploratory actions.

In this paper, we evaluate the effect of increasing camera baselines on depth perception in robot-assisted surgery. Restricted by the diameter of the surgical trocar through which they are inserted, current clinical stereo endoscopes have a fixed baseline of 5.5 mm. To overcome this restriction, we propose using a stereoscopic "pickup" camera with a side-firing design that allows for larger baselines. We conducted a user study with baselines of 10 mm, 15 mm, 20 mm, and 30 mm to evaluate the effect of increasing baseline on depth perception when used with the da Vinci surgical system. Subjects (N=28) were recruited and asked to rank differently sized poles, mounted at a distance of 200 mm from the cameras, according to their increasing order of height when viewed under different baseline conditions. The results showed that subjects performed better as the baseline was increased with the best performance at a 20 mm baseline. This preliminary proof-of-concept study shows that there is opportunity to improve depth perception in robot-assisted surgical systems with a change in endoscope design philosophy. In this paper, we present this change with our side-firing "pickup" camera and its flexible baseline design. Ultimately, this serves as the first step towards an adaptive baseline camera design that maximizes depth perception in surgery.

The human-robot collaboration is a promising and challenging field of robotics research. One of the main collaboration tasks is the object co-manipulation where the human and robot are in a continuous physical interaction and forces exerted must be handled. This involves some issues known in robotics as physical Human-Robot Interaction (pHRI), where human safety and interaction comfort are required. Moreover, a definition of interaction quality metrics would be relevant. In the current work, the assessment of Human-Robot object co-manipulation task was explored through the proposed metrics of interaction quality, based on human forces throughout the movement. This analysis is based on co-manipulation of objects with different dynamical properties (weight and inertia), with and without including these properties knowledge in the robot control law. Here, the human is a leader of task and the robot the follower without any information of the human trajectory and movement profile. For the robot control law, a well-known impedance control was applied on a 7-dof Kuka LBR iiwa 14 R820 robot. Results show that the consideration of object dynamical properties in the robot control law is crucial for a good and more comfortable interaction. Besides, human efforts are more significant with a higher no-considered weight, whereas it remains stable when these weights were considered.

Deep Reinforcement Learning has been used to exploit specific environments, but has difficulty transferring learned policies to new situations. This issue poses a problem for practical applications of Reinforcement Learning, as real-world scenarios may introduce unexpected differences that drastically reduce policy performance. We propose the use of differentiated sub-policies governed by a hierarchical controller to support adaptation in such scenarios. We also introduce a confidence- based training process for the hierarchical controller which improves training stability and convergence times. We evaluate these methods in a new Capture the Flag environment designed to explore adaptation in autonomous multi-agent settings.

Accurate uncertainty estimation associated with the pose transformation between two 3D point clouds is critical for autonomous navigation, grasping, and data fusion. Iterative closest point (ICP) is widely used to estimate the transformation between point cloud pairs by iteratively performing data association and motion estimation. Despite its success and popularity, ICP is effectively a deterministic algorithm, and attempts to reformulate it in a probabilistic manner generally do not capture all sources of uncertainty, such as data association errors and sensor noise. This leads to overconfident transformation estimates, potentially compromising the robustness of systems relying on them. In this paper we propose a novel method to estimate pose uncertainty in ICP with a Markov Chain Monte Carlo (MCMC) algorithm. Our method combines recent developments in optimization for scalable Bayesian sampling such as stochastic gradient Langevin dynamics (SGLD) to infer a full posterior distribution of the pose transformation between two point clouds. We evaluate our method, called Bayesian ICP, in experiments using 3D Kinect data demonstrating that our method is capable of both quickly and accuractely estimating pose uncertainty, taking into account data association uncertainty as reflected by the shape of the objects.

Neuromorphic hardware offers computing platforms for the efficient implementation of spiking neural networks (SNNs) that can be used for robot control. Here, we present such an SNN on a neuromorphic chip that solves a number of tasks related to simultaneous localization and mapping (SLAM): forming a map of an unknown environment and, at the same time, estimating the robot's pose. In particular, we present an SNN mechanism to detect and estimate errors when the robot revisits a known landmark and updates both the map and the path integration speed to reduce the error. The whole system is fully realized in a neuromorphic device, showing the feasibility of a purely SNN-based SLAM, which could be efficiently implemented in a small form-factor neuromorphic chip.

This paper presents a novel episodic method to learn a robot's nonlinear dynamics model and an increasingly optimal control sequence for a set of tasks. The method is based on the Koopman operator approach to nonlinear dynamical systems analysis, which models the flow of observables in a function space, rather than a flow in a state space. Practically, this method estimates a nonlinear diffeomorphism that lifts the dynamics to a higher dimensional space where they are linear. Efficient Model Predictive Control methods can then be applied to the lifted model. This approach allows for real time implementation in on-board hardware, with rigorous incorporation of both input and state constraints during learning. We demonstrate the method in a real-time implementation of fast multirotor landing, where the nonlinear ground effect is learned and used to improve landing speed and quality.

Robots often require a model of their environment to make informed decisions. In unknown environments, the ability to infer the value of a data field from a limited number of samples is essential to many robotics applications. In this work, we propose a neural network architecture to model these spatially correlated data fields based on a limited number of spatially continuous samples. Additionally, we provide a method based on biased loss functions to suggest future areas of exploration to minimize reconstruction error. We run simulated robotic information gathering trials on both the MNIST hand written digits dataset and a Regional Ocean Modeling System (ROMS) ocean dataset for ocean monitoring. Our method outperforms Gaussian process regression in both environments for modeling the data field and action selection.

Creating maps is an essential task in robotics and provides the basis for effective planning and navigation. In this paper, we learn a compact and continuous implicit surface map of an environment from a stream of range data with known poses. For this, we create and incrementally adjust an ensemble of approximate Gaussian process (GP) experts which are each responsible for a different part of the map. Instead of inserting all arriving data into the GP models, we greedily trade-off between model complexity and prediction error. Our algorithm therefore uses less resources on areas with few geometric features and more where the environment is rich in variety. We evaluate our approach on synthetic and real-world data sets and analyze sensitivity to parameters and measurement noise. The results show that we can learn compact and accurate implicit surface models under different conditions, with a performance comparable to or better than that of exact GP regression with subsampled data.

Roboticists are grappling with how to address privacy in robot design at a time when regulatory frameworks around the world increasingly require systems to be engineered to preserve and protect privacy. This paper surveys the top robotics journals and conferences over the past four decades to identify contributions with respect to privacy in robot design. Our survey revealed that less than half of one percent of the ~89,120 papers in our study even mention the word privacy. Herein, we propose privacy preserving approaches for roboticists to employ in robot design, including, assessing a robot's purpose and environment; ensuring privacy by design by selecting sensors that do not collect information that is not essential to the core objectives of that robot; embracing both privacy and performance as fundamental design challenges to be addressed early in the robot lifecycle; and performing privacy impact assessments.

In this paper, enhancement to the novel grasp planning algorithm based on gripper workspace spheres is presented. Our development requires a registered point cloud of the target from different views, assuming no prior knowledge of the object, nor any of its properties. This work features a new set of metrics for grasp pose candidates evaluation, as well as exploring the impact of high object sampling on grasp success rates. In addition to gripper position sampling, we now perform orientation sampling about the x, y, and z-axes, hence the grasping algorithm no longer require object orientation estimation. Successful experiments have been conducted on a simple jaw gripper (Franka Panda gripper) as well as a complex, high Degree of Freedom (DoF) hand (Allegro hand) as a proof of its versatility. Higher grasp success rates of 76% and 85.5% respectively has been reported by real world experiments.

In this paper, we consider a two-player racing game, where an autonomous ego vehicle has to be controlled to race against an opponent vehicle, which is either autonomous or human-driven. The approach to control the ego vehicle is based on a Sensitivity-ENhanced NAsh equilibrium seeking (SENNA) method, which uses an iterated best response algorithm in order to optimize for a trajectory in a two-car racing game. This method exploits the interactions between the ego and the opponent vehicle that take place through a collision avoidance constraint. This game-theoretic control method hinges on the ego vehicle having an accurate model and correct knowledge of the state of the opponent vehicle. However, when an accurate model for the opponent vehicle is not available, or the estimation of its state is corrupted by noise, the performance of the approach might be compromised. For this reason, we augment the SENNA algorithm by enforcing Permissive RObust SafeTy (PROST) conditions using control barrier functions. The objective is to successfully overtake or to remain in the front of the opponent vehicle, even when the information about the latter is not fully available. The successful synergy between SENNA and PROST-antithetical to the notable rivalry between the two namesake Formula 1 drivers-is demonstrated through extensive simulated experiments.

Coral species detection underwater is a challenging problem. There are many cases when even the experts (marine biologists) fail to recognize corals, hence limiting ground truth annotation for training a robust detection system. Identifying coral species is fundamental for enabling the monitoring of coral reefs, a task currently performed by humans, which can be automated with the use of underwater robots. By employing temporal cues using a tracker on a high confidence prediction by a convolutional neural network-based object detector, we augment the collected dataset for the retraining of the object detector. However, using trackers to extract examples also introduces hard or mislabelled samples, which is counterproductive and will deteriorate the performance of the detector. In this work, we show that employing a simple deep neural network to filter out hard or mislabelled samples can help regulate sample extraction. We empirically evaluate our approach in a coral object dataset, collected via an Autonomous Underwater Vehicle (AUV) and human divers, that shows the benefit of incorporating extracted examples obtained from tracking. This work also demonstrates how controlling sample generation by tracking using a simple deep neural network can further improve an object detector.

Time-optimal trajectories for dynamic robotic vehicles are difficult to compute even for state-of-the-art nonlinear programming (NLP) solvers, due to nonlinearity and bang-bang control structure. This paper presents a bilevel optimization framework that addresses these problems by decomposing the spatial and temporal variables into a hierarchical optimization. Specifically, the original problem is divided into an inner layer, which computes a time-optimal velocity profile along a given geometric path, and an outer layer, which refines the geometric path by a Quasi-Newton method. The inner optimization is convex and efficiently solved by interior-point methods. The gradients of the outer layer can be analytically obtained using sensitivity analysis of parametric optimization problems. A novel contribution is to introduce a duality gap in the inner optimization rather than solving it to optimality; this lets the optimizer realize warm-starting of the interior-point method, avoids non-smoothness of the outer cost function caused by active inequality constraint switching. Like prior bilevel frameworks, this method is guaranteed to return a feasible solution at any time, but converges faster than gap-free bilevel optimization. Numerical experiments on a drone model with velocity and acceleration limits show that the proposed method performs faster and more robustly than gap-free bilevel optimization and general NLP solvers.

Controlling and manning robots from a remote location is difficult because of the limitations one faces in perception and available degrees of actuation. Although humans can become skilled teleoperators, the amount of training time required to acquire such skills is typically very high. In this paper, we propose a novel solution (named Autocomplete) to aid novice teleoperators in manning robots adroitly. At the input side, Autocomplete relies on machine learning to detect and categorize human inputs as one from a group of motion primitives. Once a desired motion is recognized, at the actuation side an automated command replaces the human input in performing the desired action. So far, Autocomplete can recognize and synthesize lines, arcs, full circles, 3-D helices, and sine trajectories. Autocomplete was tested in simulation on the teleoperation of an unmanned aerial vehicle, and results demonstrate the advantages of the proposed solution versus manual steering.