P18241: Autonomous People Mover V/public/
Customer Handoff & Final Project Documentation
Table of Contents
Team Vision for Final Demo and Handoff
- Implement a navigation algorithm utilizing an artificial potential field method
- Modify the stop zone parameters to allow for safe stopping distance for obstacles directly ahead
- Design a demo scenario allowing for continuous driving and obstacle avoidance in a confined space
- Utilize final, fully trained computer vision model
- Tighten brake actuator to achieve faster emergency braking
- Repair brake sensor for more accurate readsings or investigate new sensors that would perform the fuction better
- Create a document for future phases, including thorough guides and video links for a helpful transition. The goal should be to allow a future team to gain a quick understanding of the APM, in order to become familiar with its operation
Test Results Summary
Inputs & Source
- Updated Test Plan
SummaryThe highest priority test plans were obstacle stopping distance, obstacle avoidance distance and smooth software controlled navigation. The specifications for a safe stopping distance were not met. This means that while the cart does stop before hitting an obstacle, the brake actuator is not quick enough to stop the cart before getting too close. The cart did not collide with any obstacles during the demonstration, but the final distance between the stopped cart and the obstacle was smaller than desired. The software controlled navigation system was implemented using an artificial potential field method. This meant that any obstacle visible to the cart created a force acting on the cart restricting its movement in that direction. With enough obstacles this meant that the cart would seek out areas with the most free space. This also means that the cart can be guided through a testing enviornment simply by walking alongside it and repelling it from driving in the wrong direction.
Risk and Problem Tracking
- Combined Risk and Problem Tracking document.
Final Project Documentation
- Final technical paper. Contains comprehensive information on all systems of the project, and how the contributions of phase five fit in with the achievements of previous phases.
- Final poster. The poster provides information at a glance about phase five's contributions to the project. The customer requirements and engineering requirements are also displayed on the poster.
- Camera mount schematic. This schematic corresponds to the custom 3D printed mount used to restrict the movement of the wide angle camera.
- Software Flow Chart. This provides an overview of how sensor data is processed and how steering/throttle commands are generated by the system.
- Image annotation instructions. This document outlines the steps required to create new image annotations for safezone training/testing.
- Caffe and ENet install instructions. This document provides step-by-step instructions for setting up the required Caffe environment needed to run safezone.
- Operating Manual. This document contains instructions for starting the cart and information on basic operation in the three modes.
Plans for Wrap-up
- Navigation: Currently, the navigation system utilizes a custom local planner, with no global planner. An Artificial Potential Field method is used, in which every obstacle contributes a force on the cart. For wandering purposes, goal points were manually added to keep the cart generally moving forward. Once goal points are decided on a global scale, the manually added local goals should be removed. For obstacle force calculations, the force applied to the cart by each obstacle has an inverse linear relationship to the distance between cart and obstacle. This causes closer obstacles to have higher forces on the cart, which can result in the cart not reacting to obstacles that are farther out. This effect can be undesirable, since the cart should ideally react to obstacles as soon as it sees them, not just if they're close.
Suggested Future WorkThroughout working on the APM, several issues were brought up that future groups should look at resolving and implementing:
- Speed control
- Currently, the APM does not have reliable speed control, so speed is completely open loop with only the existence of how fast the cart is being told to go according to throttle and remote controller throttle inputs. While there exists a set of encoder rings on the rear wheels, there is an issue with one of the encoder rings where it doesn't sense a particular spot on the ring, and thus throws off the number of measured ticks which in turn results in incorrect speed measurements. It has been brought up to the group that there exists an encoder on the drive motor which may be more reliable, so tapping into that encoder for speed measurements or purchasing a new encoder setup will drastically increase localization performance and allow for more fine grained control of the APM with knowledge of its actual speed.
- Optimizing steering, braking, and throttle
- While the steering, braking, and throttle work well, they could stand to be optimized more in terms of reaction time and force. The steering rack takes a lot of time to react in turning from one direction to the other and seems to struggle with turning in general, as indicated by strained buzzing noises from the steering components. Having faster reactive steering will allow for more fine grained control over where the cart is going.
- Attempts to tighten the braking linkage have been made, and while it has improved braking force to some extent, it still takes about 1.5 seconds to fully stop from full autonomous speed. Further exploring improving the braking force, as well as granular control of the brake, would improve safety and reaction time with respect to stopping for obstacles.
- Currently the throttle commands go through one of the Arduinos in the main pc cabinet, which leads to some jitteriness when trying to drive at low speeds. Additionally, while wandering mode does feature dynamic speed control, it is not as reactive in all situations, such as when making hard turns. Exploring an alternate method of controlling the throttle or improving the existing throttle control would help to improve overall control over the APM.
- Optimizing Artificial Potential Field for local
- The Artificial Potential Field method of local navigation based on presence of objects works well, but suffers from some issues. When multiple sensors detect the same object, or if objects such as bushes are detected, this causes a large density of object detections to occur in a small location in the occupancy grid, which in turn puts a large force vector on the cart, which can cause it to react or turn too much, potentially forcing it into situations where it will stop even though there is room for it to proceed. The Artificial Potential Field method is also limited in how far ahead we can react to objects for turning or stopping purposes. Adding more complex logic to better handle detection of objects and their influence on the APM and its direction and velocity of travel, looking further ahead to react to objects, and fine tuning of the stopping mechanics would provide huge benefits to local navigation of the APM and should be a primary objective of future teams.