ICRA 2018 Papers

We develop a multi-step motion prediction modeling method for dynamic systems over long horizons using deep learning. Building on previous work, we propose a novel hybrid network architecture, by combining deep recurrent neural networks with a quadrotor motion model created using classic system identification methods. The proposed model takes only the initial system state and motor speeds over the prediction horizon as inputs and returns robust state predictions for up to two seconds of motion at 100 Hz. We employ recurrent neural network state initialization during training, to exploit real-world dataset collected from quadrotor vehicle flights in an indoor flight arena. Our experiments demonstrate that the proposed hybrid network model consistently outperforms both black box and rigid body dynamics predictions over single and multi-step prediction scenarios, with an order of magnitude improvements in velocity estimates in particular.

We present a novel method of measurement covariance estimation that models measurement uncertainty as a function of the measurement itself. Existing work in predictive sensor modeling outperforms conventional fixed models, but requires domain knowledge of the sensors that heavily influences the accuracy and the computational cost of the models. In this work, we introduce Deep Inference for Covariance Estimation (DICE), which utilizes a deep neural network to predict the covariance of a sensor measurement from raw sensor data. We show that given pairs of raw sensor measurement and ground-truth measurement error, we can learn a representation of the measurement model via supervised regression on the prediction performance of the model, eliminating the need for hand-coded features and parametric forms. Our approach is sensor-agnostic, and we demonstrate improved covariance prediction on both simulated and real data.

Imitation learning is a powerful paradigm for robot skill acquisition. However, obtaining demonstrations suitable for learning a policy that maps from raw pixels to actions can be challenging. In this paper we describe how consumer-grade Virtual Reality headsets and hand tracking hardware can be used to naturally teleoperate robots to perform complex tasks. We also describe how imitation learning can learn deep neural network policies (mapping from pixels to actions) that can acquire the demonstrated skills. Our experiments showcase the effectiveness of our approach for learning visuomotor skills.

Robot-assisted dressing offers an opportunity to benefit the lives of many people with disabilities, such as some older adults. However, robots currently lack common sense about the physical implications of their actions on people. The physical implications of dressing are complicated by non-rigid garments, which can result in a robot indirectly applying high forces to a person's body. We present a deep recurrent model that, when given a proposed action by the robot, predicts the forces a garment will apply to a person's body. We also show that a robot can provide better dressing assistance by using this model with model predictive control. The predictions made by our model only use haptic and kinematic observations from the robot's end effector, which are readily attainable. Collecting training data from real world physical human-robot interaction can be time consuming, costly, and put people at risk. Instead, we train our predictive model using data collected in an entirely self-supervised fashion from a physics-based simulation. We evaluated our approach with a PR2 robot that attempted to pull a hospital gown onto the arms of 10 human participants. With a 0.2s prediction horizon, our controller succeeded at high rates and lowered applied force while navigating the garment around a persons fist and elbow without getting caught. Shorter prediction horizons resulted in significantly reduced performance with the sleeve catching on the participants' fists and elbows, demonstrating the value of our model's predictions. These behaviors of mitigating catches emerged from our deep predictive model and the controller objective function, which primarily penalizes high forces.

We consider the problems of learning forward models that map state to high-dimensional images and inverse models that map high-dimensional images to state in robotics. Specifically, we present a perceptual model for generating video frames from state with deep networks, and provide a framework for its use in tracking and prediction tasks. We show that our proposed model greatly outperforms standard deconvolutional methods and GANs for image generation, producing clear, photo-realistic images. We also develop a convolutional neural network model for state estimation and compare the result to an Extended Kalman Filter to estimate robot trajectories. We validate all models on a real robotic system.

In this paper we propose a new approach for learning perception-action couplings. We show that by collecting a suitable set of raw images and the associated movement trajectories, a deep encoder-decoder network can be trained that takes raw images as input and outputs the corresponding dynamic movement primitives. We propose suitable cost functions for training the network and describe how to calculate their gradients to enable effective training by back-propagation. We tested the proposed approach both on a synthetic dataset and on a widely used MNIST database to generate handwriting movements from raw images of digits. The calculated movements were also applied for digit writing with a real robot.

Localization is an indispensable component of a robot's autonomy stack that enables it to determine where it is in the environment, essentially making it a precursor for any action execution or planning. Although convolutional neural networks have shown promising results for visual localization, they are still grossly outperformed by state-of-the-art local feature-based techniques. In this work, we propose VLocNet, a new convolutional neural network architecture for 6-DoF global pose regression and odometry estimation from consecutive monocular images. Our multitask model incorporates hard parameter sharing, thus being compact and enabling real-time inference, in addition to being end-to-end trainable. We propose a novel loss function that utilizes auxiliary learning to leverage relative pose information during training, thereby constraining the search space to obtain consistent pose estimates. We evaluate our proposed VLocNet on indoor as well as outdoor datasets and show that even our single task model exceeds the performance of state-of-the-art deep architectures for global localization, while achieving competitive performance for visual odometry estimation. Furthermore, we present extensive experimental evaluations utilizing our proposed Geometric Consistency Loss that show the effectiveness of multitask learning and demonstrate that our model is the first deep learning technique to be on par with, and in some cases outperforms state-of-the-art SIFT-based approaches.

When a human performs a task with the assistance of wearable extra limbs, the human movement for performing the task may inadvertently disturb the position and orientation of the robot base, making it difficult for the robot to properly carry out its objective. Therefore, unlike self-standing robots, a wearable robot must not only assist the user without interfering or prohibiting the natural human movement, but also have the capability to detect and reject disturbances caused by the wearer's motion. This paper examines such a situation, where the human attempts to twist open a bottle while a pair of robotic fingers mounted on the same arm holds the bottle in place. As the human arm rotates to twist the cap, the robot and consequently the bottle would rotate in that same direction, which makes separation of the cap from the bottle almost impossible. To compensate for the human induced disturbances, a data-driven latent space impedance control method is developed such that the robot can secure the bottle and at the same time allow natural human movement to be carried out during manipulation. Simulation and experiments have demonstrated the efficacy of the latent space impedance controller to enable single-handed object manipulation with the assistance of wearable robotic fingers.

This paper presents a design for a decentralized adaptive controller that allows a team of agents to manipulate a common payload in \mathbb{R}^{2}R2\mathbb{R}^{2} or \mathbb{R}^{3}R3\mathbb{R}^{3}. The controller requires no communication between agents and requires no a priori knowledge of agent positions or payload properties. The agents can control the payload to track a reference trajectory in linear and angular velocity with center-of-mass measurements, in angular velocity using only local measurements and a common frame, and can stabilize its rotation with only local measurements. The controller is designed via a Lyapunov-style analysis and has proven stability and convergence. The controller is validated in simulation and experimentally with four robots manipulating an object in the plane.

Reinforcement learning (RL) is a suitable approach for controlling systems with unknown or time-varying dynamics. RL in principle does not require a model of the system, but before it learns an acceptable policy, it needs many unsuccessful trials, which real robots usually cannot withstand. It is well known that RL can be sped up and made safer by using models learned online. In this paper, we propose to use symbolic regression to construct compact, parsimonious models described by analytic equations, which are suitable for realtime robot control. Single node genetic programming (SNGP) is employed as a tool to automatically search for equations fitting the available data. We demonstrate the approach on two benchmark examples: a simulated mobile robot and the pendulum swing-up problem; the latter both in simulations and real-time experiments. The results show that through this approach we can find accurate models even for small batches of training data. Based on the symbolic model found, RL can control the system well.

Decentralized visual simultaneous localization and mapping (SLAM) is a powerful tool for multi-robot applications in environments where absolute positioning is not available. Being visual, it relies on cheap, lightweight and versatile cameras, and, being decentralized, it does not rely on communication to a central entity. In this work, we integrate state-of-the-art decentralized SLAM components into a new, complete decentralized visual SLAM system. To allow for data association and optimization, existing decentralized visual SLAM systems exchange the full map data among all robots, incurring large data transfers at a complexity that scales quadratically with the robot count. In contrast, our method performs efficient data association in two stages: first, a compact full-image descriptor is deterministically sent to only one robot. Then, only if the first stage succeeded, the data required for relative pose estimation is sent, again to only one robot. Thus, data association scales linearly with the robot count and uses highly compact place representations. For optimization, a state-of-the-art decentralized pose-graph optimization method is used. It exchanges a minimum amount of data which is linear with trajectory overlap. We characterize the resulting system and identify bottlenecks in its components. The system is evaluated on publicly available datasets and we provide open access to the code. Supplementary Material Data and code are at: https://github.com/uzh-rpg/dslam_open.

The goal of this paper is to present an end-to-end, data-driven framework to control Autonomous Mobility-on-Demand systems (AMoD, i.e. fleets of self-driving vehicles). We first model the AMoD system using a time-expanded network, and present a formulation that computes the optimal rebalancing strategy (i.e., preemptive repositioning) and the minimum feasible fleet size for a given travel demand. Then, we adapt this formulation to devise a Model Predictive Control (MPC) algorithm that leverages short-term demand forecasts based on historical data to compute rebalancing strategies. Using simulations based on real customer data from DiDi Chuxing, we test the end-to-end performance of this controller with a state-of-the-art LSTM neural network to predict customer demand: we show that this approach scales very well for large systems (indeed, the computational complexity of the MPC algorithm does not depend on the number of customers and of vehicles in the system) and outperforms state-of-the-art rebalancing strategies by reducing the mean customer wait time by up to to 89.6 %.

Modeling realistic human joint limits is important for applications involving physical human-robot interaction. However, setting appropriate human joint limits is challenging because it is pose-dependent: the range of joint motion varies depending on the positions of other bones. The paper introduces a new technique to accurately simulate human joint limits in physics simulation. We propose to learn an implicit equation to represent the boundary of valid human joint configurations from real human data. The function in the implicit equation is represented by a fully connected neural network whose gradients can be efficiently computed via back-propagation. Using gradients, we can efficiently enforce realistic human joint limits through constraint forces in a physics engine or as constraints in an optimization problem.

In this paper we categorise swarming into four classes, depending on the manner in which swarm members communicate. We identify two of these classes as ready candidates for the provision of communications within tactical defence networks in which radio-frequency communications are highly contested or denied. We demonstrate the feasibility of a swarm-robotics approach to data ferrying from both of these classes using simulation, emulation, and physical swarm robotic platforms within indoor flight facilities. The results show strong alignment between data dissemination capabilities of the simulated and physical systems; we envisage these techniques providing communications between not only troops, but also other swarm robotic platforms, thereby enabling swarm robotics applications and human-swarm interaction within harsh communications environments.

A recent trend in optimal motion planning has broadened the research area toward the hybridization of sampling, optimization and grid-based approaches. We can expect that synergy from such integrations leads to overall performance improvement, but seamless integration and generalization is still an open problem. In this paper, we suggest a hybrid motion planning algorithm utilizing a sampling-based and optimization-based planner while simultaneously approximating a configuration free space. Unlike conventional optimization-based approaches, the proposed algorithm does not depend on a priori information or resolution-complete factors, e.g., a distance field. Ours instead learns spatial information on the fly by exploiting empirical information during the execution, and decentralizes the information over the constructed graph for efficient access. With the help of the learned information, our optimization-based local planner exploits the local area to identify the connectivity of configuration free space without depending on the precomputed domain knowledge. To show the novelty of proposed algorithm, we evaluate it against other asymptotic optimal planners in both synthetic and complex benchmarks with varying degrees of freedom. We also discuss the performance improvement, properties and limitations we have observed.

Densely connected networks for classification enable feature exploration and result in state-of-the-art performance on multiple classification tasks. The alternative to dense networks is the residual network which enables feature re-usage. In this work, we combine these orthogonal concepts for encoder-decoder architectures, which we call Dual-Path Dense-Block Network (DPDB-Net). We introduce a dense block which incorporates feature re-usage and new feature exploration in the encoder. Moreover, we discuss that feature re-usage by the residual network architecture leads to a feature map explosion in the decoder and, thus, is not advantageous in this part of the network. We evaluated our proposed architecture in multiple segmentation tasks and report state-of-the-art performance on the Freiburg Forest dataset and competitive results on the Cam Vid dataset.

Modelling the physical properties of everyday objects is a fundamental prerequisite for autonomous robots. We present a novel generative adversarial network (DEFO-NET), able to predict body deformations under external forces from a single RGB-D image. The network is based on an invertible conditional Generative Adversarial Network (IcGAN) and is trained on a collection of different objects of interest generated by a physical finite element model simulator. Defo-netinherits the generalisation properties of GANs. This means that the network is able to reconstruct the whole 3-D appearance of the object given a single depth view of the object and to generalise to unseen object configurations. Contrary to traditional finite element methods, our approach is fast enough to be used in real-time applications. We apply the network to the problem of safe and fast navigation of mobile robots carrying payloads over different obstacles and floor materials. Experimental results in real scenarios show how a robot equipped with an RGB-D camera can use the network to predict terrain deformations under different payload configurations and use this to avoid unsafe areas.

Use of low-cost depth sensors, such as a stereo camera setup with illuminators, is of particular interest for numerous applications ranging from robotics and transportation to mixed and augmented reality. The ability to quantify noise is crucial for these applications, e.g., when the sensor is used for map generation or to develop a sensor scheduling policy in a multi-sensor setup. Range error models provide uncertainty estimates and help weigh the data correctly in instances where range measurements are taken from different vantage points or with different sensors. Such a model is derived in this work. We show that the range error for stereo systems with integrated illuminators is cubic and validate the proposed model experimentally with an off-the-shelf structured light stereo system. The experiments confirm the validity of the model and simplify the application of this type of sensor in robotics.

In recent years, many different feedback controllers for robotic applications have been proposed and implemented. However, the high coupling between the different software modules made their integration into one common architecture difficult. Consequently, this has hindered the ability of a user to employ the different controllers into a single, general and modular framework. To address this problem, we present Ctrl-MORE, a software architecture developed to fill the gap between control developers and other users in robotic applications. On one hand, Ctrl-MORE aims to provide developers with an opportunity to integrate easily and share their controllers with other roboticists working in different areas. For example, manipulation, locomotion, vision and so on. On the other hand, it provides to end-users a tool to apply the additional control strategies that guarantee the execution of desired behaviors in a transparent, yet efficient way. The proposed control architecture allows an easier integration of general purpose feedback controllers, such as stabilizers, with higher control layers such as trajectory planners, increasing the robustness of the overall system.