IROS 2019 Papers

Identifying objects using sparse tactile sensor arrays requires movement across the surface and the integration of sensory information to support hypotheses. In this study we demonstrate a surface placement control strategy to orient and position a biomimetic tactile whisker array relative to the object surface. This reduces the variance in tactile view, or representation, of each region of the object surface thus relaxing the prior pose dependence for object recall. In addition we evaluate two regional search strategies in an attempt to further improve the robustness and speed of object classification.

Mobile robots can be considered completely autonomous if they embed active algorithms for Simultaneous Localization And Mapping (SLAM). This means that the robot is able to autonomously, or actively, explore and create a reliable map of the environment, while simultaneously estimating its pose. In this paper, we propose a novel framework to robustly solve the active SLAM problem, in scenarios in which some prior information about the environment is available in the form of a topo-metric graph. This information is typically available or can be easily developed in industrial environments, but it is usually affected by uncertainties. In particular, the distinguishing features of our approach are: the inclusion of prior information for solving the active SLAM problem; the exploitation of this information to pursue active loop closure; the on-line correction of the inconsistencies in the provided data. We present some experiments, that are performed in different simulated environments: the results suggest that our method improves on state-of-the-art approaches, as it is able to deal with a wide variety of possibly large uncertainties.

Enabling robots to act according to human preferences across diverse environments is a crucial task, extensively studied by both roboticists and machine learning researchers. To achieve it, human preferences are often encoded by a reward function which the robot optimizes for. This reward function is generally static in the sense that it does not vary with time or the interactions. Unfortunately, such static reward functions do not always adequately capture human preferences, especially, in non-stationary environments: Human preferences change in response to the emergent behaviors of the other agents in the environment. In this work, we propose learning reward dynamics that can adapt in non-stationary environments with several interacting agents. We define reward dynamics as a tuple of reward functions, one for each mode of interaction, and mode-utility functions governing transitions between the modes. Reward dynamics thereby encodes not only different human preferences but also how the preferences change. Our contribution is in the way we adapt preference-based learning into a hierarchical approach that aims at learning not only reward functions but also how they evolve based on interactions. We derive a probabilistic observation model of how people will respond to the hierarchical queries. Our algorithm leverages this model to actively select hierarchical queries that will maximize the volume removed from a continuous hypothesis space of reward dynamics. We empirically demonstrate reward dynamics can match human preferences accurately.

In this work, we propose a framework to learn an inverse model of redundant systems. We address three problems. By formalizing what it actually means to learn an inverse model, we derive a method where the inverse model, represented as a neural network, is learned by minimizing an upper bound on the real performance error, which is provided by a forward model (kernel regression or Gaussian process) learned on the currently available data. Most machine learning methods focus on learning the mapping of the function. For inverse models, it is, however, crucial to know the reachable set of the true forward model, since this becomes the domain of the inverse. Therefore, we secondly propose a method to estimate the reachable set of the system. Finally, we develop an active exploration strategy that is based on maximizing a lower bound on the true fill-distance to efficiently generate the data in the high dimensional input space. A key feature of our method is that the resulting learned inverse model provides error bounds on its performance.From an application point of view, this work is motivated by learning to control musculoskeletal systems. In the experiments, we show for both a simulated model of a human arm with six muscles and a real muscle-driven robot that the proposed method is able to learn the reachable set of these systems as well as a policy that enables to accurately control the position.

Relative pose estimation is critical for collaborative multi-agent systems. To achieve accurate and low-cost localization in cluttered and GPS-denied environments, we propose a novel relative pose estimation system based on a designed active infrared coded target. Specifically, each agent is equipped with a forward-looking monocular camera and a unique infrared coded target. The target with the unique lighted LED arrangement is detected by the camera and processed with an efficient decoding algorithm. The relative pose between the agent and the camera is estimated by combining a PnP algorithm and a Kalman filter. Various experiments are performed to show that the proposed pose estimation system is accurate, robust and efficient in cluttered and GPS-denied environments.

Contextual skill models are learned to provide skills over a range of task parameters, often using regression across optimal task-specific policies. However, the sequential nature of the learning process is usually neglected. In this paper, we propose to use active incremental learning by selecting a task which maximizes performance improvement over entire task set. The proposed framework exploits knowledge of individual tasks accumulated in a database and shares it among the tasks using a contextual skill model. The framework is agnostic to the type of policy representation, skill model, and policy search. We evaluated the skill improvement rate in two tasks, ball-in-a-cup and basketball. In both, active selection of tasks lead to a consistent improvement in skill performance over a baseline.

Action recognition has wide applications in assisted living, health monitoring, surveillance, and human-computer interaction. In traditional action recognition methods, RGB video-based ones are effective but computationally inefficient, while skeleton-based ones are computationally efficient but do not make use of low-level detail information. This work considers action recognition based on a multimodal fusion between the 3D skeleton and the RGB image. We design a neural network that uses a 3D skeleton sequence and a single middle frame from an RGB video as input. Specifically, our method picks up one frame in a video and extracts spatial features from it using two attention modules, a self-attention module and a skeleton-attention module. Further, temporal features are extracted from the skeleton sequence via a BI-LSTM sub-network. Finally, the spatial features and the temporal features are combined via a feature fusion network for action classification. A distinct feature of our method is that it uses only a single RGB frame rather than an RGB video. Accordingly, it has a light-weighted architecture and is more efficient than RGB video-based methods. Comparative evaluation on two public datasets, NTU-RGBD and SYSU, demonstrates that, our method can achieve competitive performance compared with state-of-the-art methods.

Transformable aerial robots are favorable in aerial manipulation tasks for their flexible ability to change configuration during the flight. By assuming robot keeping in the mild motion, the previous researches sacrifice aerial agility to simplify the complex non-linear system into a single rigid body with a linear controller. In this paper, we present a framework towards agile swing motion for the transformable multi-links aerial robot. We introduce a computational-efficient non-linear model predictive controller and joints motion primitive frame-work to achieve agile transforming motions and validate with a novel robot named HYRURS-X. Finally, we implement our framework under a table tennis task to validate the online and agile performance.Supplementary MaterialThis paper is accompanied by a experiment video: http://www.jsk.t.u-tokyo.ac.jp/%7eshifan/paper/iros19/video.mp4.

Humans perform the task of pouring often and in which exhibit consistent accuracy regardless of the complicated dynamics of the liquid. Model predictive control (MPC) appears to be a natural candidate solution for the task of accurate pouring considering its wide use in industrial applications. However, MPC requires the model of the system in question. Since an accurate model of the liquid dynamics is difficult to obtain, the usefulness of MPC for the pouring task is uncertain. In this work, we model the dynamics of water using a recurrent neural network (RNN), which enables the use of MPC for pouring control. We evaluated our RNN-enabled MPC controller using a physical system we made ourselves and averaged a pouring error of 16.4 mL over 5 different source containers. We also compared our controller with a baseline switch controller and showed that our controller achieved a much higher accuracy than the baseline controller.

This paper aims to improve the practical scalability of automated tools to assist in designing robots. Such problems rapidly become intractable because the underlying design space is immense. We consider a specific type of design tool addressed in prior work, which constructs a representation of the destructiveness boundary in the space of robot designs. This prior work showed that a legible representation, specifically a decision tree, of this boundary can illuminate which elements of a sensing or actuation system are most important for enabling the robot to complete its task. In that context, the robot's interaction with the world is represented as procrustean graph, and the space of robot designs is represented by the space of label maps that rewrite the labels on that graph. In this paper, we expand upon those results by showing how domain knowledge can enable such tools to find solutions to more complex problems within a reasonable time frame. Specifically, we propose three different scenarios, expressed as constraints on the p-graph and on the label maps, under which the learning algorithm to identify the destructiveness boundary can converge quickly to high accuracy results for problems at larger scales than the prior, general-purpose algorithm. The conditions for each of these scenarios are easily verifiable and the set of problems that fall under each is rich enough to encompass several interesting problems. Experimental results demonstrate the effectiveness of the proposed methods.

Tightly coupled Visual-Inertial Navigation System (VINS) implementations have proven their superiority due to their ability to jointly optimize all state variables. However, this joint optimization is considered as a computational bottleneck within the system, and thus many traditional VINS can only be implemented on platforms containing powerful processors. In this work, we show a significant reduction in the computational burden of optimization can be achieved through the use of fragments; a set of co-visible feature points from across several different cameras efficiently split, categorized into groups, and then analyzed using a machine-learning inspired frequent-pattern growth algorithm. Furthermore, we use a reduced continuous representation during preintegration for better accuracy while requiring less computational resources. We validate our algorithm in datasets, showing that the derivation is not only more accurate, but requires significantly less computational resources. When tested on a Raspberry Pi, our implementation was able to track the system's state in nearly 20 frames per second using only a single CPU core. Testing on another low power module shows a power draw of around 800 mW. Due to run-time considerations the Raspberry Pi was only competitive in regards to other optimization methodologies, but our algorithm displays remarkable accuracy on a more powerful platform.

We present a neural network approach and publicly available dataset for developing and benchmarking algorithms for localization on the Moon. Accurate localization is essential for navigation, path planning, and science objectives. On Earth, localization can be achieved using a satellite navigation system; such a system is, however, unavailable for other planetary bodies. Current absolute and relative localization methods for rovers on planetary surfaces are laborious manual tasks or accumulate significant location errors over time resulting in incorrect location estimates, respectively. Recently, Earth-based approaches have explored localization by matching surface-perspective imagery to satellite imagery using computer vision and machine learning. Those focusing on natural environments have shown such image matching to be feasible for localization, though it remains challenging. The purpose of this project was to produce a working proof-of-concept model to match surface-perspective imagery to satellite images for planetary roving applications. This would improve absolute localization accuracy (relative to current methods) by utilizing machine learning algorithms to offer greater automation and faster calculations, and produce a benchmark dataset for future studies. To achieve these, the following objectives were met: (1) configuration and modification of an existing synthetic planetary environment using Unreal Engine 4 to generate surface and satellite perspective images, (2) reprojection of 2.4 + million surface perspective images to 600,000 + satellite perspective images, and (3) matching of said reprojected images to satellite images via a Siamese convolutional neural network, designed and trained for this task.

Detecting and localizing contacts acting on a manipulator is a relevant problem for manipulation tasks like grasping, since contact information can be helpful for recovering from collisions or for improving the grasping performance itself. In this work, we present a solution for contact point localization, which is based on Monte Carlo Localization. Usually, an Articulated Robotic Manipulator (ARM) is not equipped with tactile skin, but with proprioceptive sensors, which we assume as an input for our method. In our experiments, we compare our method with a direct optimization method, machine learning approaches and another particle filter method, both on simulated and real world data from a Kinova Jaco2. While our proposed method clearly outperforms the other optimization approaches, it performs about equally well as Random Forest (RF) classifiers, although both methods have their strengths on different parts of the manipulator, and even achieves better results than multi-layer perceptions (MLPs) on the links farthest from the manipulator base.

Trajectory optimization is a widely used tool for robot motion planning and control. Existing solvers for these problems either rely on off-the-shelf nonlinear programming solvers that are numerically robust and capable of handling arbitrary constraints, but tend to be slow because they are general purpose; or they use custom numerical methods that take advantage of the problem structure to be fast, but often lack robustness and have limited or no ability to reason about constraints. This paper presents ALTRO (Augmented Lagrangian TRajectory optimizer), a solver for constrained trajectory optimization problems that handles general nonlinear state and input constraints and offers fast convergence and numerical robustness thanks to careful exploitation of problem structure. We demonstrate its performance on a set of benchmark motion-planning problems and offer comparisons to the standard direct collocation method with large-scale sequential quadratic programming and interior-point solvers.

Measuring forces and torques applied to tools of unconventional shapes can be challenging due to difficulties in connecting commercially available force/torque sensors. We propose an innovative, customizable six-axis force/torque sensor which can be clamped around a tool or end-effector to minimize the proximity between tool tip and sensor structure. The sensor is fabricated using 3d printing technology and consists of two pieces which together form an 8-legged Stewart platform. Each leg is designed as a cantilever beam to allow for a measurable displacement under an external load. The displacements of the legs are measured with 8 light intensity-based optoelectronic sensors, which exhibit high sensitivities and low noise levels without the need for external amplification circuitry. A customized printed circuit board and peripheral hardware are proposed to allow for efficient analog-to-digital conversion of the force/torque measurement. A calibration process is proposed which makes use ofa commercial ATI Mini40 sensor and custom hardware to allow for fast calibration routines. Finally, the calibrated sensor design is compared to the ATI Mini40 sensor by measuring sequences of forces and torques, and the maximum errors of force/torque components (Fx, Fy, Fz, Mx, My, Mz) are 16.3%, 20.0%, 27.5%, 20.5%, 21.6%, 14.9% respectively.

We study self-supervised adaptation of a robot's policy for social interaction, i.e., a policy for active communication with surrounding pedestrians through audio or visual signals. Inspired by the observation that humans continually adapt their behavior when interacting under varying social context, we propose Adaptive EXP4 (A-EXP4), a novel online learning algorithm for adapting the robot-pedestrian interaction policy. To address limitations of bandit algorithms in adaptation to unseen and highly dynamic scenarios, we employ a mixture model over the policy parameter space. Specifically, a Dirichlet Process Gaussian Mixture Model (DPMM) is used to cluster the parameters of sampled policies and maintain a mixture model over the clusters, hence effectively discovering policies that are suitable to the current environmental context in an unsupervised manner. Our simulated and real-world experiments demonstrate the feasibility of A-EXP4 in accommodating interaction with different types of pedestrians while jointly minimizing social disruption through the adaptation process. While the A-EXP4 formulation is kept general for application in a variety of domains requiring continual adaptation of a robot's policy, we specifically evaluate the performance of our algorithm using a suitcase-inspired assistive robotic platform. In this concrete assistive scenario, the algorithm observes how audio signals produced by the navigational system affect the behavior of pedestrians and adapts accordingly. Consequently, we find A-EXP4 to effectively adapt the interaction policy for gently clearing a navigation path in crowded settings, resulting in significant reduction in empirical regret compared to the EXP4 baseline.

We introduce Pressure Field Contact (PFC), an approximate model for predicting the contact surface, pressure distribution, and net contact wrench between nominally rigid objects. PFC combines and generalizes two ideas: a bed of springs (an `elastic foundation') and hydrostatic pressure. Continuous pressure fields are computed offline for the interior of each nominally rigid object. Unlike hydrostatics or elastic foundations, the pressure fields need not satisfy mechanical equilibrium conditions. When two objects nominally overlap, a contact surface is defined where the two pressure fields are equal. This static pressure is supplemented with a dissipative rate-dependent pressure and friction to determine tractions on the contact surface. The contact wrench between pairs of objects is an integral of traction contributions over this surface. PFC evaluates much faster than elasticity-theory models, while showing the essential trends of force, moment, and stiffness increase with contact load. It yields continuous wrenches even for non-convex objects and coarse meshes. The method shows promise as sufficiently fast, accurate, physical, and robust for robotics applications including motion and tactile sensor simulation, controller learning and synthesis, state estimation, and design-in-simulation.

Haptic simulators are used in many education domains to improve hands-on training performance. Many of these simulators are used in the medical domain, where it is particularly important to be trained, guided by an expert. However, the majority of haptic hands-on training devices are usable by only one user at the same time, either the trainer or the trainee. Dual-user approaches have enhanced this situation by allowing both trainer and trainee to use the same simulator to introduce guided training into the simulation. In this paper, we introduce a multi-trainee architecture that expands further a former dual-user hands-on haptic training solution designed according to an energetic approach. This system will permit several trainees to be involved in the same simulation at the same time, guided by a single trainer. It paves the way to the multi-trainee hands-on training on haptic simulators.

Performance evaluation of urban autonomous vehicles (AVs) requires a realistic model of the behavior of other road users in the environment. Learning such models from data involves collecting naturalistic data of real-world human behavior. In many cases, acquisition of this data can be prohibitively expensive or intrusive. Additionally, the available data often contain only typical behaviors and exclude behaviors that are classified as rare events. To evaluate the performance of AVs in such situations, we develop a model of traffic behavior based on the theory of bounded rationality. Based on the experiments performed on a large naturalistic driving data, we show that the developed model can be applied to estimate probability of rare events, as well as to generate new traffic situations for testing.