IROS 2017 论文列表

There are two major challenges in universal stair climbing: stairs without riser and with nose, and stairs with various dimensions. In this study, we proposed an indoor robot platform to overcome these challenges. First, to create an angle of attack, the Tusk, a passive, protruded element, was added in front of a 4-wheel robot. For design analysis and optimization of the Tusk, a simplified model of universal stair climbing was applied. To accommodate stairs without risers and with nose, the assistive track mechanism was applied. To climb the stair regardless of its dimension, length-adjustable mechanism was added. The results indicated the robot with these mechanisms successfully overcame each challenge. The performance was better than most conventional stair-climbing robots in terms of the range of compatible stairs. We expect these new approaches to expand the range of indoor robot operation with minimal cost.

Car racing control is often assumed to take place in isolation, although the greatest strategic challenges arise in the presence of other cars. Our solution uses a combined nonlinear system model which includes the four-wheeled vehicle dynamics model and point-mass model of its opponent car. This proposed combined model allows the optimal control scheme to automatically find an overtaking strategy. The FORCES code generation tool to handle the consequent computational complexity of the Nonlinear Model Predictive Controller (NMPC). We demonstrate the practicality of the suggested method by performing real-time control in a racing simulation software.

Adaptive robot hands have changed the way we approach and think of robot grasping and manipulation. Traditionally, pinch, fingertip grasping and dexterous, in-hand manipulation tasks were executed with fully actuated, rigid robot hands and relied on analytic methods, computation of the hand object Jacobians and extensive numerical simulations for deriving optimal and minimal effort grasps. However, even insignificant uncertainties in the modeling space could render the extraction of candidate grasps or manipulation paths infeasible. Adaptive hands use underactuated mechanisms and structural compliance, facilitating by design the successful extraction of stable grasps and the robust execution of manipulation tasks, even under significant object pose or other environmental uncertainties. In this paper, we propose a methodology for the automated extraction of dexterous, in-hand manipulation strategies / primitives for adaptive hands. To do so, we use a constrained optimization scheme that describes the kinematics of adaptive hands during the grasping and manipulation processes, an automated experimental setup for data collection, a clustering technique that groups together similar manipulation strategies, and a dimensionality reduction technique that projects the robot kinematics to lower dimensional manifolds. In these manifolds, control is simplified and hand operation becomes more intuitive. In this work, we also assess the effect of the extracted manipulation primitives on the object pose perturbations. The efficiency of the proposed methods is experimentally verified for various adaptive robot hands. The extracted primitives can simplify the operation and control of the open-source robot hand designs of the Yale Open Hand project in dexterous manipulation tasks.

Most recent approaches to 3D pose estimation from RGB-D images address the problem in a two-stage pipeline. First, they learn a classifier-typically a random forest-to predict the position of each input pixel on the object surface. These estimates are then used to define an energy function that is minimized w.r.t. the object pose. In this paper, we focus on the first stage of the problem and propose a novel classifier based on a depth-aware Convolutional Neural Network. This classifier is able to learn a scale-adaptive regression model that yields very accurate pixel-level predictions, allowing to finally estimate the pose using a simple RANSAC-based scheme, with no need to optimize complex ad hoc energy functions. Our experiments on publicly available datasets show that our approach achieves remarkable improvements over state-of-the-art methods.

Six degree-of-freedom (DoF) pose feedback is essential for the development of closed-loop control techniques for microrobotics. This paper presents two methods for depth estimation of transparent microrobots inside an Optical Tweezers (OT) setup using image sharpness measurements and model-based tracking. The x-y position and the 3D orientation of the object are estimated using online model-based template matching. The proposed depth estimation methodologies are validated experimentally by comparing the results with the ground truth.

There have been increasing demands for developing robotic system combining camera and inertial measurement unit in navigation task, due to their low-cost, lightweight and complementary properties. In this paper, we present a Visual Inertial Odometry (VIO) system which can utilize sparse depth to estimate 6D pose in GPS-denied and unstructured environments. The system is based on Multi-State Constraint Kalman Filter (MSCKF), which benefits from low computation load when compared to optimization-based method, especially on resource-constrained platform. Features are enhanced with depth information forming 3D landmark position measurements in space, which reduces uncertainty of position estimate. And we derivate measurement model to access compatibility with both 2D and 3D measurements. In experiments, we evaluate the performance of the system in different in-flight scenarios, both cluttered room and industry environment. The results suggest that the estimator is consistent, substantially improves the accuracy compared with original monocular-based MSKCF and achieves competitive accuracy with other research.

The paper exploits weak Manhattan constraints to parse the structure of indoor environments from RGB-D video sequences in an online setting. We extend the previous approach for single view parsing of indoor scenes to video sequences and formulate the problem of recovering the floor plan of the environment as an optimal labeling problem solved using dynamic programming. The temporal continuity is enforced in a recursive setting, where labeling from previous frames is used as a prior term in the objective function. In addition to recovery of piecewise planar weak Manhattan structure of the extended environment, the orthogonality constraints are also exploited by visual odometry and pose graph optimization. This yields reliable estimates in the presence of large motions and absence of distinctive features to track. We evaluate our method on several challenging indoors sequences demonstrating accurate SLAM and dense mapping of low texture environments. On existing TUM benchmark [19] we achieve competitive results with the alternative approaches which fail in our environments.

While dense visual SLAM methods are capable of estimating dense reconstructions of the environment, they suffer from a lack of robustness in their tracking step, especially when the optimisation is poorly initialised. Sparse visual SLAM systems have attained high levels of accuracy and robustness through the inclusion of inertial measurements in a tightly-coupled fusion. Inspired by this performance, we propose the first tightly-coupled dense RGB-D-inertial SLAM system. Our system has real-time capability while running on a GPU. It jointly optimises for the camera pose, velocity, IMU biases and gravity direction while building up a globally consistent, fully dense surfel-based 3D reconstruction of the environment. Through a series of experiments on both synthetic and real world datasets, we show that our dense visual-inertial SLAM system is more robust to fast motions and periods of low texture and low geometric variation than a related RGB-D-only SLAM system.

Movement Primitives (MPs) have been widely used over the last years for learning robot motion tasks with direct Policy Search (PS) reinforcement learning. Among them, Probabilistic Movement Primitives (ProMPs) are a kind of MP based on a stochastic representation over sets of trajectories, which benefits from the properties of probability operations. However, the generation of such ProMPs requires a set of demonstrations to capture motion variability. Additionally, using context variables to modify trajectories coded as MPs is a popular approach nowadays in order to adapt motion to environmental variables. This paper proposes a contextual representation of ProMPs that allows for an easy adaptation to changing situations through context variables, by reparametrizing motion with them. Moreover, we propose a way of initializing contextual trajectories without the need of real robot demonstrations, by setting an initial position, a final position, and a number of trajectory interest points, where the contextual variables are evaluated. The parametrizations obtained show to be accurate while relieving the user from the need of performing costly computations such as conditioning. Additionally, using this contextual representation, we propose a simple yet effective quadratic optimization-based obstacle avoidance method for ProMPs. Experiments in simulation and on a real robot show the promise of the approach.

In autonomous Mobility on Demand (MOD) systems, customers request rides from a fleet of shared vehicles that can be automatically positioned in response to customer demand. Recent approaches to MOD systems have focused on environments where customers can only request rides through an app or by waiting at a station. This paper develops MOD fleet management approaches for ride hailing, where customers may instead request rides simply by hailing a passing vehicle, an approach of particular importance for campus MOD systems. The challenge for ride hailing is that customer demand is not explicitly provided as it would be with an app, but rather customers are only served if a vehicle happens to be located at the arrival location. This work focuses on maximizing the number of served hailing customers in an MOD system by learning and utilizing customer demand. A Bayesian framework is used to define a novel customer demand model which incorporates observed pedestrian traffic to estimate customer arrival locations with a quantification of uncertainty. An exploration planner is proposed which routes MOD vehicles in order to reduce arrival rate uncertainty. A robust ride hailing fleet management planner is proposed which routes vehicles under the presence of uncertainty using a chance-constrained formulation. Simulation of a real-world MOD system on MIT's campus demonstrates the effectiveness of the planners. The customer demand model and exploration planner are demonstrated to reduce estimation error over time and the ride hailing planner is shown to improve the fraction of served customers in the system by 73% over a baseline exploration approach.

Search-based planning that uses a state lattice has been successfully applied in many applications but its utility is limited when confronted with complex problems represented by a lattice with many nodes and edges with high branching factor. However, in many seemingly complex problems, proper “form-fitting” can reduce the number of nodes and edges needed to represent the problems, provides a concise state lattice and therefore simplifies the computation. This paper proposes a planning framework which strikes to identify concise representations of problems, creates such “form-fitting” state lattice on which a more concise search can take place. In a sense, we take a conventional state lattice and map it onto a deformed space, and then the motion primitives and heuristics follow. Since the contribution of the paper is not the search approach but rather the means by which the lattice is deformed, any search-based planner can then be easily changed to a corresponding deformed version with no increase in time complexity. This paper demonstrate the benifits of the approach which includes 1) planned path can be followed with few changes in motion primitives and thus can provide global smoothness of planned path; 2) fewer states are expanded and thus shorter time to search solution in state space is required, and 3) fewer states are expanded and thus less memory is required to save the state lattice. We demonstrate the benefit of the proposed approach in illustrative toy examples, as well as robot experiments.

This paper demonstrates a feasible method for using a deep neural network as a sensor to estimate the attitude of a flying vehicle using only flight video. A dataset of still images and associated gravity vectors was collected and used to perform supervised learning. The network builds on a previously trained network and was trained to be able to approximate the attitude of the camera with an average error of about 8 degrees. Flight test video was recorded and processed with a relatively simple visual odometry method. The aircraft attitude is then estimated with the visual odometry as the state propagation and network providing the attitude measurement in an extended Kalman filter. Results show that the proposed method of having the neural network provide a gravity vector attitude measurement from the flight imagery reduces the standard deviation of the attitude error by approximately 12 times compared to a baseline approach.

Robots are expected to operate autonomously in dynamic environments. Understanding the underlying dynamic characteristics of objects is a key enabler for achieving this goal. In this paper, we propose a method for pointwise semantic classification of 3D LiDAR data into three classes: non-movable, movable and dynamic. We concentrate on understanding these specific semantics because they characterize important information required for an autonomous system. To learn the distinction between movable and non-movable points in the environment, we introduce an approach based on deep neural network and for detecting the dynamic points, we estimate pointwise motion. We propose a Bayes filter framework for combining the learned semantic cues with the motion cues to infer the required semantic classification. In extensive experiments, we compare our approach with other methods on a standard benchmark dataset and report competitive results in comparison to the existing state-of-the-art. Furthermore, we show an improvement in the classification of points by combining the semantic cues retrieved from the neural network with the motion cues.

In this paper we consider the problem of robot navigation in simple maze-like environments where the robot has to rely on its onboard sensors to perform the navigation task. In particular, we are interested in solutions to this problem that do not require localization, mapping or planning. Additionally, we require that our solution can quickly adapt to new situations (e.g., changing navigation goals and environments). To meet these criteria we frame this problem as a sequence of related reinforcement learning tasks. We propose a successor-feature-based deep reinforcement learning algorithm that can learn to transfer knowledge from previously mastered navigation tasks to new problem instances. Our algorithm substantially decreases the required learning time after the first task instance has been solved, which makes it easily adaptable to changing environments. We validate our method in both simulated and real robot experiments with a Robotino and compare it to a set of baseline methods including classical planning-based navigation.

The high precision assembly of mechanical parts requires precision that exceeds that of robots. Conventional part-mating methods used in the current manufacturing require numerous parameters to be tediously tuned before deployment. We show how a robot can successfully perform a peg-in-hole task with a tight clearance through training a recurrent neural network with reinforcement learning. In addition to reducing manual effort, the proposed method also shows a better fitting performance with a tighter clearance and robustness against positional and angular errors for the peg-in-hole task. The neural network learns to take the optimal action by observing the sensors of a robot to estimate the system state. The advantages of our proposed method are validated experimentally on a 7-axis articulated robot arm.

Precise localization of robots is imperative for their safe and autonomous navigation in both indoor and outdoor environments. In outdoor scenarios, the environment typically undergoes significant perceptual changes and requires robust methods for accurate localization. Monocular camera-based approaches provide an inexpensive solution to such challenging problems compared to 3D LiDAR-based methods. Recently, approaches have leveraged deep convolutional neural networks (CNNs) to perform place recognition and they turn out to outperform traditional handcrafted features under challenging perceptual conditions. In this paper, we propose an approach for directly regressing a 6-DoF camera pose using CNNs and a single monocular RGB image. We leverage the idea of transfer learning for training our network as this technique has shown to perform better when the number of training samples are not very high. Furthermore, we propose novel data augmentation in 3D space for additional pose coverage which leads to more accurate localization. In contrast to the traditional visual metric localization approaches, our resulting map size is constant with respect to the database. During localization, our approach has a constant time complexity of O(1) and is independent of the database size and runs in real-time at ~80 Hz using a single GPU. We show the localization accuracy of our approach on publicly available datasets and that it outperforms CNN-based state-of-the-art methods.

Skilled robot task learning is best implemented by predictive action policies due to the inherent latency of sensorimotor processes. However, training such predictive policies is challenging as it involves finding a trajectory of motor activations for the full duration of the action. We propose a data-efficient deep predictive policy training (DPPT) framework with a deep neural network policy architecture which maps an image observation to a sequence of motor activations. The architecture consists of three sub-networks referred to as the perception, policy and behavior super-layers. The perception and behavior super-layers force an abstraction of visual and motor data trained with synthetic and simulated training samples, respectively. The policy super-layer is a small subnetwork with fewer parameters that maps data in-between the abstracted manifolds. It is trained for each task using methods for policy search reinforcement learning. We demonstrate the suitability of the proposed architecture and learning framework by training predictive policies for skilled object grasping and ball throwing on a PR2 robot. The effectiveness of the method is illustrated by the fact that these tasks are trained using only about 180 real robot attempts with qualitative terminal rewards.

In this paper, we develop a novel 3D object recognition algorithm to perform detection and pose estimation jointly. We focus on analyzing the advantages of the 3D point cloud relative to the RGB-D image and try to eliminate the unpredictability of output values that inevitably occurs in regression tasks. To achieve this, we first adopt the Truncated Signed Distance Function (TSDF) to encode the point cloud and extract low compact discriminative feature via unsupervised deep learning network. This approach can not only eliminate the dense scale sampling for offline model training but also reduce the distortion by mapping the 3D shape to the 2D plane and overcome the dependence on color cues. Then, we train a Hough forests to achieve multi-object detection and 6-DoF pose estimation simultaneously. In addition, we propose a robust multilevel verification strategy that effectively reduces the unpredictability of output values which occurs in the hough regression module. Experiments on public datasets demonstrate that our approach provides effective results comparable to the state-of-the-arts.

Reliable lane detection is a fundamental necessity for driver assistance, driver safety functions and fully automated vehicles. Based on other detection and classification tasks, deep learning based methods are likely to yield the most accurate outputs for detecting lane markers, but require vast amounts of labeled data. We propose to train a deep neural network for detecting lane markers in camera images without manually labeling any images. To achieve this, we project high definition maps for automated driving into our image and correct for misalignments due to inaccuracies in localization and coordinate frame transformations. The corrections are performed by calculating the offset between features within our map and detected ones in the images. By using detections in the actual image for refining the projections, our labels become close to pixel perfect. After a fast, visual quality check, our projected lane markers can be used for training a fully convolutional network to segment lane markers in images. A single worker can easily generate 20,000 of those labels within a single day. Our fully convolutional network is trained only on automatically generated labels. All of our detections are based solely on gray-scale mono camera inputs without any additional information. The resulting network regularly detects clean lane markers at distances of around 150 meters on a 1 Megapixel camera.