No description available.

Operation of Autonomous Mobile Robots (AMRs) of all forms that include wheeled ground vehicles, quadrupeds and humanoids in dynamically changing GPS denied environments without a-priori maps, exclusively using onboard sensors, is an unsolved problem that has potential to transform the economy, and vastly improve humanity's capabilities with improvements to agriculture, manufacturing, disaster response, military and space exploration. Conventional AMR automation approaches are modularized into perception, motion planning and control which is computationally inefficient, and requires explicit feature extraction and engineering, that inhibits generalization, and deployment at scale. Few works have focused on real-world end-to-end approaches that directly map sensor inputs to control outputs due to the large amount of well curated training data required for supervised Deep Learning (DL) which is time consuming and labor intensive to collect and label, and sample inefficiency and challenges to bridging the simulation to reality gap using Deep Reinforcement Learning (DRL). This dissertation presents a novel method to efficiently train DRL with significantly fewer samples in a constrained racetrack environment at physical limits in simulation, transferred zero-shot to the real-world for robust end-to-end AMR navigation. The representation learned in a compact parameter space with 2 fully connected layers with 64 nodes each is demonstrated to exhibit emergent behavior for Out-of-Distribution (OOD) generalization to navigation in new environments that include unstructured terrain without maps, dynamic obstacle avoidance, and navigation to objects of interest with vision input that encompass low light scenarios with the addition of a night vision camera. The learned policy outperforms conventional navigation algorithms while consuming a fraction of the computation resources, enabling execution on a range of AMR forms with varying embedded computer payloads. / Doctor of Philosophy / Robots with wheels or legs to move around environments improve humanity's capabilities in many applications such as agriculture, manufacturing, and space exploration. Reliable, robust mobile robots have the potential to significantly improve the economy. A key component of mobility is navigation to either explore the surrounding environment, or travel to a goal position or object of interest by avoiding stationary, and dynamic obstacles. This is a complex problem that has no reliable solution, which is one of the main reasons robots are not present everywhere, assisting people in various tasks. Past and current approaches involve first mapping an environment, then planning a collision-free path, and finally executing motor signals to traverse along the path. This has several limitations due to the lack of detailed pre-made maps, and inability to operate in previously unseen, dynamic environments. Furthermore, these modular methods require high computation resources due to the large number of calculations required for each step that prevents high real-time speed, and functionality in small robots with limited weight capacity for onboard computers, that are beneficial for reconnaissance, and exploration tasks. This dissertation presents a novel Artificial Intelligence (AI) method for robot navigation that is more computationally efficient than current approaches, with better performance. The AI model is trained to race in simulation at multiple times real-time speed for cost-effective, accelerated training, and transferred to a physical mobile robot where it retains its training experience, and generalizes to navigation in new environments without maps, with exploratory behavior, and dynamic obstacle avoidance capabilities.

This thesis project is part of an internship at Visual Behavior. The company aims at producing computer vision models for robotics, helping the machine to better understand the world through the camera eye. The image holds many features that deep learning models are able to extract: navigable area, depth inference and object detection. Example of recent advances are the RAFTstereo model [1] to infer or refine depth features from stereo images, or the end-to-end Object detection model DETR [2]. The field of autonomous navigation can then benefit from these advanced features to propose better path planning methods. In particular, to help the deployment of ground robots in human crowded environments, the robots behavior must not only be safe but it must also look smart so as to inspire trust. This thesis proposes a local path planner based on the Dynamic Window Approach [3] using a scoring function inspired from the Velocity Obstacle method [4] so as to benefit from the flexibility of the first and the long-term anticipation of the second. The proposed method can induce a smart behavior by setting the robot on safe tracks from a long time horizon without increasing the time to reach a positional goal, compared to a closer-ranged strategy inspired from the DW4DO method [5]. This improves the robot’s ability to deal with several moving obstacles and to avoid engaging in already occupied corridors. The code produced in this thesis uses ROS and the gazebo simulator and is available in the following git page https://github.com/FloCoic oi/fc_thesis along with the minimal instructions to run the install and get started to quickly run a demo. / Detta examensarbete är en del av en praktik på Visual Behavior. Företaget har som mål att ta fram modeller för datorseende för robotar som hjälper maskinen att bättre förstå världen genom kamerans öga. Bilden innehåller många egenskaper som modeller för djupinlärning kan extrahera: navigerbart område, djupinferens och objektsdetektering. Exempel på nya framsteg är RAFT-stereo-modellen [1] för att härleda eller förfina djupegenskaper från stereobilder, eller ”end-to-end” objektdetektering modellen DETR [2]. Inom området autonom navigering kan man sedan dra nytta av dessa avancerade funktioner för att föreslå bättre metoder för vägplanering. För att underlätta användningen av markrobotar i miljöer med mycket människor måste robotarnas beteende inte bara vara säkert utan också se smart ut så att de väcker förtroende. I den här avhandlingen föreslås en lokal vägplanerare som bygger på Dynamic Window Approach [3] och som använder en poängfunktion inspirerad av Velocity Obstacle metoden [4] för att dra nytta av flexibiliteten hos den första metoden och den långsiktiga förutsebarheten hos den andra. Den föreslagna metoden kan framkalla ett smart beteende genom att sätta roboten på säkra spår på lång sikt utan att öka tiden för att nå ett positionsmål, jämfört med en strategi med närmare avstånd som inspirerats av DW4DOmetoden [5]. Detta förbättrar robotens förmåga att hantera flera rörliga hinder och att undvika att gå in i redan upptagna korridorer. Koden som produceras i denna avhandling använder ROS och gazebosimulatorn och finns tillgänglig på följande git-sida https://github.c om/FloCoicoi/fc_thesis tillsammans med minimala instruktioner för att köra installationen och komma igång för att snabbt köra en demo.