ICRA 2018 Papers

Developing advanced robots can be a resource-intensive activity that creates many challenges for research teams. There is a need to formulate new techniques for systematically designing complex robot systems, especially in cases where high adaptability is needed and design metrics cannot be explicitly specified in advance. This research explores how the Goals/Questions/Metrics (GQM) method, a well-established technique for process measurement, can be modified for use as a design tool in robotics. To illustrate how a design-orientated GQM method may be used in practice, a sample use-case is given detailing how the approach was applied to the task of developing a bespoke robotic gripper for a service robot. The study provides an early indication that the adoption of GQM principles by designers can have significant benefits in robotics applications. However, further investigation is needed to better understand the magnitude and scope of any improvements.

Achieving both the flexibility and robustness required to advance the use of robotics in small and medium-sized productions is an essential but difficult task. A fundamental problem is making the robot run blindly without additional sensors while still being robust to uncertainties and variations in the assembly processes. In this paper, we address the use of parameterized motions suitable for blind execution and robust to uncertainties in the assembly process. Collisions and incorrect assemblies are detected based on robot motor currents while motion parameters are updated based on Bayesian Optimization utilizing Gaussian Process learning. This allows for motion parameters to be optimized using real world trials which incorporate all uncertainties inherent in the assembly process without requiring advanced robot and sensor setups. The result is a simple and straightforward system which helps the user automatically find robust and uncertainty-tolerant motions. We present experiments for an assembly case showing both detection and learning in the real world and how these combine to a robust robot system.

Learning from demonstration algorithms, such as Inverse Reinforcement Learning, aim to provide a natural mechanism for programming robots, but can often require a prohibitive number of demonstrations to capture important subtleties of a task. Rather than requesting additional demonstrations blindly, active learning methods leverage uncertainty to query the user for action labels at states with high expected information gain. However, this approach can still require a large number of labels to adequately reduce uncertainty and may also be unintuitive, as users are not accustomed to determining optimal actions in a single out-of-context state. To address these shortcomings, we propose a novel trajectory-based active Bayesian inverse reinforcement learning algorithm that (1) queries the user for critiques of automatically generated trajectories, rather than asking for demonstrations or action labels, (2) utilizes trajectory segmentation to expedite the critique / labeling process, and (3) predicts the user's critiques to generate the most highly informative trajectory queries. We evaluated our algorithm in simulated domains, finding it to compare favorably to prior work and a randomized baseline.

In this paper, we consider motion as a means of sending messages between robots. We focus on a scenario in which a message is encoded in a sending robot's trajectory, and decoded by a receiver robot equipped with a monocular camera. The relative pose between the robots is unknown. We introduce an online Bayesian estimation algorithm based on the Multi-hypothesis Extended Kalman Filter for the receiving robot to simultaneously estimate its relative pose to the sender, and the trajectory class of the sender. The difficulty in this problem arises from the monocular vision model of the receiver and the unknown relative pose between robots, which brings inherent ambiguity into the trajectory identification, and hence the message decoding. An active vision-based control policy is derived and combined with the Bayesian estimation in order to deal with this difficulty. The policy is constructed online based on Monte Carlo Tree Search and aims at reducing the entropy over the trajectory class distribution. The algorithm has broad applications, e.g., to intent modeling and motion prediction for autonomous driving and autonomous drone operations. Simulation results demonstrate that the proposed estimation algorithm and the control policy result in an accurate trajectory classification.

Image-based modeling techniques [1]–[3] can now generate photo-realistic 3D models from images. But it is up to users to provide high quality images with good coverage and view overlap, which makes the data capturing process tedious and time consuming. We seek to automate data capturing for image-based modeling. The core of our system is an iterative linear method to solve the multi-view stereo (MVS) problem quickly and plan the Next-Best-View (NBV) effectively. Our fast MVS algorithm enables online model reconstruction and quality assessment to determine the NBVs on the fly. We test our system with a toy unmanned aerial vehicle (UAV) in simulated, indoor and outdoor experiments. Results show that our system improves the efficiency of data acquisition and ensures the completeness of the final model.

Robotic assistance allows surgeons to perform dexterous and tremor-free procedures, but is still underrepresented in deep brain neurosurgery and endonasal surgery where the workspace is constrained. In these conditions, the vision of surgeons is restricted to areas near the surgical tool tips, which increases the risk of unexpected collisions between the shafts of the instruments and their surroundings, in particular in areas outside the surgical field-of-view. Active constraints can be used to prevent the tools from entering restricted zones and thus avoid collisions. In this paper, a vector field inequality is proposed that guarantees that tools do not enter restricted zones. Moreover, in contrast with early techniques, the proposed method limits the tool approach velocity in the direction of the forbidden zone boundary, guaranteeing a smooth behavior and that tangential velocities will not be disturbed. The proposed method is evaluated in simulations featuring two eight degrees-of-freedom manipulators that were custom-designed for deep neurosurgery. The results show that both manipulator-manipulator and manipulator-boundary collisions can be avoided using the vector field inequalities.

Humans represent and discriminate the objects in the same category using their properties, and an intelligent robot should be able to do the same. In this paper, we build a robot system that can autonomously perceive the object properties through touch. We work on the common object category of clothing. The robot moves under the guidance of an external Kinect sensor, and squeezes the clothes with a GelSight tactile sensor, then it recognizes the 11 properties of the clothing according to the tactile data. Those properties include the physical properties, like thickness, fuzziness, softness and durability, and semantic properties, like wearing season and preferred washing methods. We collect a dataset of 153 varied pieces of clothes, and conduct 6616 robot exploring iterations on them. To extract the useful information from the high-dimensional sensory output, we applied Convolutional Neural Networks (CNN) on the tactile data for recognizing the clothing properties, and on the Kinect depth images for selecting exploration locations. Experiments show that using the trained neural networks, the robot can autonomously explore the unknown clothes and learn their properties. This work proposes a new framework for active tactile perception system with vision-touch system, and has potential to enable robots to help humans with varied clothing related housework.

In this paper we present a system for the in situ33In the context of building construction, “in situ” means that fabrication takes place at the structure's final location directly on the building site. fabrication of a full-scale, load-bearing, and doubly-curved steel reinforced concrete wall. Two complementary vision-based sensing systems provide the feedback necessary to build a 12 meter long steel wire mesh as part of a novel digital building process. The sensing systems provide estimates of the robot pose, referenced to the CAD model of the building site, as well as feedback on the accuracy of the built structure over the course of construction. This second piece of information is used to adapt the building plan to compensate for system inaccuracies and material deformations which occur during buildup. In this way, the structure was successfully built with 98% of the total geometry within 2 centimeters of the designed position. To the best of our knowledge, this is the largest structure which has been built by a mobile robot using solely vision-based sensing.

This paper deals with a new accelerated path integral method, which iteratively searches optimal controls with a small number of iterations. This study is based on the recent observations that a path integral method for reinforcement learning can be interpreted as gradient descent. This observation also applies to an iterative path integral method for optimal control, which sets a convincing argument for utilizing various optimization methods for gradient descent, such as momentum-based acceleration, step-size adaptation and their combination. We introduce these types of methods to the path integral and demonstrate that momentum-based methods, like Nesterov Accelerated Gradient and Adam, can significantly improve the convergence rate to search for optimal controls in simulated control systems. We also demonstrate that the accelerated path integral could improve the performance on model predictive control for various vehicle navigation tasks. Finally, we represent this accelerated path integral method as a recurrent network, which is the accelerated version of the previously proposed path integral networks (PI-Net). We can train the accelerated PI-Net more efficiently for inverse optimal control with less RAM than the original PI-Net.

Online learning of a robot's inverse dynamics model for trajectory tracking necessitates an interaction between the robot and its environment to collect training data. This is challenging for physical robots in the real world, especially for humanoids and manipulators due to their large and high dimensional state and action spaces, as a large amount of data must be collected over time. This can put the robot in danger when learning tabula rasa and can also be a time-intensive process especially in a multi-robot setting, where each robot is learning its model from scratch. We propose accelerating learning of the inverse dynamics model for trajectory tracking tasks in this multi-robot setting using knowledge transfer, where robots share and re-use data collected by preexisting robots, in order to speed up learning for new robots. We propose a scheme for collecting a sample of correspondences from the robots for training transfer models, and demonstrate, in simulations, the benefit of knowledge transfer in accelerating online learning of the inverse dynamics model between several robots, including between a low-cost Interbotix PhantomX Pincher arm, and a more expensive and relatively heavier Kuka youBot arm. We show that knowledge transfer can save up to 63% of training time of the youBot arm compared to learning from scratch, and about 58% for the lighter Pincher arm.

In this paper, we investigate the use of surrogate agents to accelerate test scenario generation for autonomous vehicles. Our goal is to train the surrogate to replicate the true performance modes of the system. We create these surrogates by utilizing imitation learning with deep neural networks. By using imitator surrogates in place of the true agent, we are capable of predicting mission performance more quickly, gaining greater throughput for simulation-based testing. We demonstrate that using on-line imitation learning with Dataset Aggregation (DAgger) can not only correctly encode a policy that executes a complex mission, but can also encode multiple different behavioral modes. To improve performance for the target vehicle and mission, we manipulate the training set during each iteration to remove samples which do not contribute to the final policy. We call this approach Quantile-DAgger (Q-DAgger) and demonstrate its ability to replicate the behaviors of an autonomous vehicle in a collision avoidance scenario.

We present a pneumatic actuator capable of changing length by 1000%, applying both pushing and pulling forces, and independently modulating its length and stiffness. These characteristics are enabled by individually addressable internal and external chambers that work antagonistically against one another. The high deformation with low hysteresis is achieved by wrinkling of thin materials that are assumed to be inextensible but flexible, as opposed to stretchable. A model for the actuator is presented and validated with experimental results, showing capabilities of high strain, pushing and pulling, and independent control of length and stiffness. These characteristics are motivated by the application of a compliant truss robot. Accordingly, we show a simple grounded tetrahedron with three actuator elements and three static elements. We demonstrate motion of the tetrahedron apex against external loads and the ability of the structure to vary its stiffness. The actuator offers a unique set of characteristics that could increase the capabilities of soft robotic devices.

Drones are enabling new forms of cinematography. Aerial filming via drones in action scenes is difficult because it requires users to understand the dynamic scenarios and operate the drone and camera simultaneously. Existing systems allow the user to manually specify the shots and guide the drone to capture footage, while none of them employ aesthetic objectives to automate aerial filming in action scenes. Meanwhile, these drone cinematography systems depend on the external motion capture systems to perceive the human action, which is limited to the indoor environment. In this paper, we propose an Autonomous CinemaTography system “ACT” on the drone platform to address the above the challenges. To our knowledge, this is the first drone camera system which can autonomously capture cinematic shots of action scenes based on limb movements in both indoor and outdoor environments. Our system includes the following novelties. First, we propose an efficient method to extract 3D skeleton points via a stereo camera. Second, we design a real-time dynamical camera planning strategy that fulfills the aesthetic objectives for filming and respects the physical limits of a drone. At the system level, we integrate cameras and GPUs into the limited space of a drone and demonstrate the feasibility of running the entire cinematography system onboard in real-time. Experimental results in both simulation and real-world scenarios demonstrate that our cinematography system “ACT” can capture more expressive video footage of human action than that of a state-of-the-art drone camera system.

Iterative Closest Point (ICP) is a widely used method for performing scan-matching and registration. Being simple and robust, this method is still computationally expensive and may be challenging to use in real-time applications with limited resources on mobile platforms. In this paper we propose a novel effective method for acceleration of ICP which does not require substantial modifications to the existing code. This method is based on an idea of Anderson acceleration which is an iterative procedure for finding a fixed point of contractive mapping. The latter is often faster than a standard Picard iteration, usually used in ICP implementations. We show that ICP, being a fixed point problem, can be significantly accelerated by this method enhanced by heuristics to improve overall robustness. We implement proposed approach into Point Cloud Library (PCL) and make it available online. Benchmarking on the real-world data fully supports our claims.

This work presents a new attitude balancing strategy implemented and validated on a quadrupedal robot equipped with a custom hybrid wheel-legged mobility system. The proposed method uses an inverse kinematics solution scheme based on Quadratic Programming optimization to generate full body motions that ensure the desired balancing performances. The strategy generates a compliant behaviour to cope with the applied external forces resulting in a stable and smooth reaction response. Furthermore, the method takes advantage of the robot hybrid wheeled-legged mobility system to provide new motion capabilities and balancing reactions as it will be shown through the paper. Extensive simulation studies on the Centauro robot are presented. Results show the efficiency of the propose method demonstrating significant contribution in the rejection of the applied external disturbances.

This paper addresses the problem of gait estimation using visual and inertial data, as well as human motion models. Specifically, a batch least-squares (BLS) algorithm is presented that fuses data from a minimal set of sensors [two inertial measurement units (IMUs), one on each foot, and a head-mounted IMU-camera pair] along with motion constraints corresponding to the different walking states, to estimate the person's head and feet poses. Subsequently, gait models are employed to solve for the lower-body's posture and generate its animation. Experimental results against the VICON motion capture system demonstrate the accuracy of the proposed minimal sensors-based system for determining a person's motion.

Our Tensegrity-Inspired Compliant Three degree-of-freedom (DOF) robotic joint adds omnidirectional compliance to robotic limbs while reducing sprung mass through base mounted actuation. This enables a robotic limb which is safer to operate alongside humans and fragile equipment while still capable of generating quick movements and large forces if required. Unlike many other soft robotic systems which leverage continuously soft materials, our joint is simpler to model with low order dynamic systems and has a host of embedded sensing which provide ample information of its position and velocity. We first discuss geometry selection and optimization to maximize the theoretical configuration space of the joint. We then show several of our mechatronic design solutions, which are easily generalized to a multitude of cable-driven mechanisms, and demonstrate the performance of these mechanisms within the context of our hardware prototype. We then present results on the controllable stiffness of our physical prototype. Finally, we demonstrate the strength of our prototype which is capable of lifting a 7 kg mass at a distance of 0.95 meters from the joint.

The ability of rope-climbing robots in aloft operation is limited by its self-supporting and locomotion ability. In many applications, a given task is also too complex to be achieved by a single rope-climbing robot acting alone. The solution of multiple rope-climbing robots can overcome the limitations. However, existing control methods for rope-climbing robots are limited to single robot, and the open issue of coordination between multiple rope-climbing robots has not been systematically addressed. This paper presents a new synchronization scheme for position control of multiple rope-climbing robots, such that each robot moves to the corresponding desired position while synchronizing the heights between each other. Maintaining the same height is very important to guarantee the stability of the task-oriented manipulator installed among multiple robots, when it is performing the manipulation task. The development of the proposed controller is based on the singular perturbation approach, by treating the fast actuator dynamics as a perturbation of the slow robot dynamics, such that the lowest control complexity is achieved. The exponential stability of the overall system that consists of the fast and slow subsystems is proved by using Tikhonov’ s theorem. Experimental results are presented to illustrate the performance of the proposed controller.

Voxel volumes are simple to implement and lend themselves to many of the tools and algorithms available for 2D images. However, the additional dimension of voxels may be costly to manage in memory when mapping large spaces at high resolutions. While lowering the resolution and using interpolation is common work-around, in the literature we often find that authors either use trilinear interpolation or nearest neighbors and rarely any of the intermediate options. This paper presents a survey of geometric interpolation methods for voxel-based map representations. In particular we study the truncated signed distance field (TSDF) and the impact of using fewer than 8 samples to perform interpolation within a depth-camera pose tracking and mapping scenario. We find that lowering the number of samples fetched to perform the interpolation results in performance similar to the commonly used trilinear interpolation method, but leads to higher frame-rates. We also report that lower bit-depth generally leads to performance degradation, though not as much as may be expected, with voxels containing as few as 3 bits sometimes resulting in adequate estimation of camera trajectories.