[Paper Review] AutOTranS: an Autonomous Open World Transportation System

Brayan S. Zapata-Impata, Vikrant Shah+2 more|arXiv (Cornell University)|Oct 8, 2018

Robotic Path Planning Algorithms13 references4 citations

TL;DR

AutOTranS proposes an autonomous mobile manipulation system for outdoor transportation of novel objects in unstructured environments. It integrates navigation, grasp detection via GPD, and a task planner to achieve 85.7% overall task success in picking up, transporting, and dropping trash bags, garbage, tools, and fruits using only user-defined pick and drop points.

ABSTRACT

Tasks in outdoor open world environments are now ripe for automation with mobile manipulators. The dynamic, unstructured and unknown environments associated with such tasks -- a prime example would be collecting roadside trash -- makes them particularly challenging. In this paper we present an approach to solving the problem of picking up, transporting, and dropping off novel objects outdoors. Our solution integrates a navigation system, a grasp detection and planning system, and a custom task planner. We perform experiments that demonstrate that the system can be used to transport a wide class of novel objects (trash bags, general garbage, gardening tools and fruits) in unstructured settings outdoors with a relatively high end-to-end success rate of 85%. See it at work at: https://youtu.be/93nWXhaGEWA

Motivation & Objective

• Address the challenge of autonomously transporting novel, unstructured outdoor objects in dynamic, unstructured environments.
• Enable mobile robots to perform end-to-end pick-and-place tasks without prior object knowledge or environmental modifications.
• Develop a system that operates with minimal human input, requiring only initial pick and drop point selection on a map.
• Overcome limitations of prior systems that require known objects, visual tags, or human teleoperation.
• Demonstrate robustness across diverse outdoor scenarios involving trash, tools, and organic materials.

Proposed method

• Utilizes a mobile base (Warthog), UR10 robotic arm with Robotiq 85 gripper, Intel RealSense D415 cameras, and SICK LMS511 laser scanner for perception and mobility.
• Employs a custom task planner that supports two transport strategies: 'collect all' and 'collect one by one' to manage object transport sequences.
• Applies the Grasp Pipeline Detector (GPD) for grasp detection without prior object models, enabling grasp planning on novel, unsegmented objects.
• Uses 2D laser scanning for robot localization and navigation, with point cloud registration via fixed cameras or hand-eye camera setup.
• Integrates a multi-stage process: point cloud registration, grasp calculation, grasp execution, and navigation to target points.
• Employs a map-based navigation system that allows autonomous path planning to user-defined pick and drop locations in previously explored areas.

Experimental results

Research questions

• RQ1 Can a mobile manipulator autonomously navigate to and from user-defined outdoor locations without prior environmental modeling?
• RQ2 Can a grasp detection system successfully identify and execute grasps on completely novel, unstructured outdoor objects without object-specific training?
• RQ3 What is the end-to-end success rate of transporting diverse object types (e.g., trash bags, fruits, tools) in real-world outdoor environments?
• RQ4 How do navigation and grasping performance vary across different object classes and transport strategies?
• RQ5 What are the primary failure modes in open-world outdoor manipulation, and how can they be mitigated?

Key findings

• The system achieved an overall task success rate of 85.7% across four experimental scenarios involving trash bags, general garbage, gardening tools, and fruits.
• Navigation success was 96.1%, with the robot successfully reaching target points in most trials despite environmental variability.
• Grasping success rates varied by object type: 80.8% for all objects combined, with 90.9% success for fruits on the ground and 76.9% when retrieved from a basket.
• The 'collect all' strategy outperformed 'collect one by one' in terms of task completion, particularly for objects like fruits and tools.
• The primary failure modes were weak grasps due to thin handles (e.g., trash bag handles) and object slips during lifting, especially with long or fragile items.
• Execution time averaged 118 seconds per task (min: 44s, max: 239s), with navigation being the longest phase, averaging 98.18s for pick and 137.86s for drop.