Topological SLAM | EPFL Graph Search

This thesis is about topological navigation, more precisely about space representation, perception, localization and mapping. All these elements are needed in order to obtain a robust and reliable framework for navigation. This is essential in order to move in an environment, manipulate objects in it and avoid collisions. The method proposed in this dissertation is suitable for fully autonomous mobile robots, operating in structured indoor and outdoor environments. High robustness is necessary to obtain a distinctive and reliable representation of the environment. This can be obtained by combining the information acquired by several sensors with complementary characteristics. A multimodal perception system that includes the wheel encoders for odometry and two extereoceptive sensors composed of a laser range finder that gives a 360° view of the environment and an omnidirectional camera are used in this work. Significant, robust and stable features are extracted from sensory data. The different features extracted from the extereoceptive sensors (i.e. corners, vertical edges, color patches, etc.) are fused and combined into a single, circular, and distinctive space representation named the fingerprint of a place. This representation is adapted for topological navigation and is the foundation for the whole dissertation. One of the major tasks of a mobile robot is localization. Different topological localization approaches based on the fingerprint concept are presented in this dissertation. Localization on a fingerprint-based representation is reduced to a problem of fingerprint matching. Two of these methods make use of the Bayesian Programming (BP) formalism and two others are based on dynamic programming. They also show how multimodal perception increases the reliability of topological localization for mobile robots. In order to autonomously acquire and create maps, robots have to explore their environment. Several exploration tools for indoor environments are presented: wall following, mid-line following, center of free space of a room, door detection, and environment structure identification. An automatic and incremental topological mapping system based on fingerprints of places and a global localizer using Partially Observable Markov Decision Processes (POMDP) are proposed. The construction of a topological mapping system is combined with localization, both relying on fingerprints of places, in order to perform Simultaneous Localization and Mapping (SLAM). This enables navigation of an autonomous mobile robot in a structured environment without relying on maps given a priori, without using artificial landmarks and by employing a semantic spatial representation that allows a more natural interface between humans and robots. The fingerprint approach, combining the information from all sensors available to the robot, reduces perceptual aliasing and improves the distinctiveness of places. This fingerprint-based approach yields a consistent and distinctive representation of the environment and is extensible in that it permits spatial cognition beyond just pure navigation. All these methodologies have been validated through experiments. Indoor and outdoor experiments have been conducted over a distance exceeding 2 km. The fingerprints of places proved to provide a compact and distinctive methodology for space representation and place recognition – they permit encoding of a huge amount of placerelated information in a single circular sequence of features. The experiments have verified the efficacy and reliability of this approach.

On-line Independent Support Vector Machines for Cognitive Systems

Barbara Caputo, Jie Luo

Learning from experience and adapting to changing stimuli are fundamental capabilities for artificial cognitive systems. This calls for on-line learning methods able to achieve high accuracy while at the same time using limited computer power. Research on autonomous agents has been actively investigating these issues, mostly using probabilistic frameworks and within the context of navigation and learning by imitation. Still, recent results on robot localization have clearly pointed out the potential of discriminative classifiers for cognitive systems. In this paper we follow this approach and propose an on-line version of the Support Vector Machine (SVM) algorithm. Our method, that we call On-line Independent SVM, builds a solution on-line, achieving an excellent accuracy vs.~compactness trade-off. In particular the size of the obtained solution is always bounded, implying a bounded testing time. At the same time, the algorithm converges to the optimal solution at each incremental step, as opposed to similar approaches where optimality is achieved in the limit of infinite number of training data. These statements are supported by experiments on standard benchmark databases as well as on two real-world applications, namely

(a)

place recognition by a mobile robot in an indoor environment, and

(b)

human grasping posture classification.

IDIAP2007

A Holistic Approach to Energy Harvesting for Indoor Robots

Florian Christopher Vaussard

Service robotics is a fast expanding market. Inside households, domestic robots can now accomplish numerous tasks, such as floor cleaning, surveillance, or remote presence. Their sales have considerably increased over the past years. Whereas 1.05 million domestic service robots were reportedly sold in 2009, at least 2.7 million units were sold in 2013. Consequently, this growth gives rise to an increase of the energy needs to power such a large and growing fleet of robots. However, the unique properties of mobile robots open some new fields of research. We must find technologies that are suitable for decreasing the energy requirements and thus further advance towards a sustainable development. This thesis tackles two fundamental goals based on a holistic approach of the global problem. The first goal is to reduce the energy needs by identifying key technologies in making energy-efficient robots. The second goal is to leverage innovative indoor energy sources to increase the ratio of renewable energies scavenged from the environment. To achieve our first goal, new energy-wise metrics are applied to real robotic hardware. This gives us the means to assess the impact of some technologies on the overall energy balance. First, we analysed seven robotic vacuum cleaners from a representative sample of the market that encompasses a wide variety of technologies. Simultaneous Localisation and Mapping (SLAM) was identified as a key technology to reduce energy needs when carrying out such tasks. Even if the instantaneous power is slightly increased, the completion time of the task is greatly reduced. We also analysed the needed sensors to achieve SLAM, as they are largely diversified. This work tested three sensors using three different technologies. We identified several important metrics. As of our second goal, potential energy sources are compared to the needs of an indoor robot. The sunshine coming through a building's apertures is identified as a promising source of renewable power. Numerical simulations showed how a mobile robot is mandatory to take full advantage of this previously unseen situation, as well as the influence of the geometric parameters on the yearly energy income under ideal sunny conditions. When considering a real system, the major difficulty to overcome is the tracking of the sunbeam along the day. The proposed algorithm uses a hybrid method. A high-level cognitive approach is responsible for the initial placement. Following realignments during the day are performed by a low-level reactive behaviour. A solar harvesting module was developed for our research robot. The tests conducted inside a controlled environment demonstrate the feasibility of this concept and the good performances of the aforementioned algorithm. Based on a realistic scenario and weather conditions, we computed that between 1 and 14 days of recharge could be necessary for a single cleaning task. In the future, our innovative technology could greatly lower the energy needs of service robots. However, it is not completely possible to abandon the recharge station due to occasional bad weather. The acceptance of this technology inside the user's home ecosystem remains to be studied.

EPFL2015

Real-Time Wide-Baseline Place Recognition Using Depth Completion

Zetao Chen

Place recognition is an essential capability for robotic autonomy. While ground robots observe the world from generally similar viewpoints over repeated visits, other robots, such as small aircraft, experience far more different viewpoints, requiring place recognition for images captured from very wide baselines. While traditional feature-based methods fail dramatically under extreme viewpoint changes, deep learning approaches demand heavy runtime processing. Driven by the need for cheaper alternatives able to run on computationally restricted platforms, such as small aircraft, this work proposes a novel real-time pipeline employing depth-completion on sparse feature maps that are anyway computed during robot localization and mapping, to enable place recognition at extreme viewpoint changes. The proposed approach demonstrates unprecedented precision-recall rates on challenging benchmarking and own synthetic and real datasets with up to 45◦ difference in viewpoints. In particular, our synthetic datasets are, to the best of our knowledge, the first to isolate the challenge of viewpoint changes for place recognition, addressing a crucial gap in the literature. All of the new datasets are publicly available to aid benchmarking.

2019