BBR Progress Report 006: Autonomous 2-D Mapping of a Building Floor

Transcription

1 BBR Progress Report 006: Autonomous 2-D Mapping of a Building Floor Andy Sayler & Constantin Berzan November 30, 2010 Abstract In the past two weeks, we implemented and tested landmark extraction based on the RANSAC algorithm. We implemented the prediction phase of the EKF, and part of the correction phase. We have found inconsistencies in the tutorial we are following, which we are now trying to reconcile. The next step is to finish implementing the EKF, either by computing the equations ourselves, or by finding another source. The code for this project is located at 1 Introduction SLAM is the problem of Simultaneous Localization and Mapping using a mobile robot. We aim to develop a proof-of-concept SLAM system adhering to the behavior-based philosophy. The robot will use odometry and laser range data to navigate the first floor of Halligan, and produce an image representation of the floor map. 2 Project Summary We will be developing our robot using a schema (or possibly hybrid) architecture. The robot will utilize schema based functions to manifest the following behaviors: Avoid obstacles Avoid local minima Seek new areas Utilize SLAM (laser + odometry) to deduce current location Utilize SLAM (laser + odometry) to generate persistent environmental map We will utilize the ADE robotics environment to complete our implementation. The first objective will be to create code capable of navigating our robot through unfamiliar environments and exploring these environments to their full potential without getting stuck or colliding with obstacles. 1

2 The next goal will be to build a map of the environment and provide localization abilities. This will be done by implementing a basic SLAM system using laser and odomtery data. We aim to develop a 2D floor plan map using data from our SLAM system. Should we complete the initial scope of this project ahead of schedule, we may opt to pursue one or more of the following extensions: Utilize vision data in SLAM system Utilize vision data in 2D map generation Compare performance of our SLAM system to Carmen Augment SLAM system with additional sensor packages (radio ranging, etc) 3 Problems tackled Literature review: SLAM for Dummies tutorial both Carmen Andy Vision-based SLAM Constantin Discussed and refined project scope: Opted to implement SLAM ourselves Decided to use SLAM data as primary source for persistent 2D map data Logistical issues: Decided on workflow and code layout for interfacing our code with ADE Figured out bootstrapping of ADE registry and simulator Wrote SSH scripts for easy remote operation of the robot Discussed Odometry Testing Setup: Linear Distance Testing (i.e. Drive 10m, then measure actual distance and angle) Start/Stop Testing (i.e. drive 1m in 10cm steps issuing start and stop commands between each step to measure start/stop drift) Rotational Distance Testing (i.e. Turn 180 degrees, then measure actual angle.) Acceleration Test (i.e. drive 10m and use internal timer to log distance at each tenth second interval to compute robot dynamics) Created Map of Halligan for the Simulator: The map was auto-generated by extracting all rectangles from a SVG image, which was drawn by hand The svg-to-xml conversion script is reusable, and allows the map to be modified easily, using a graphical editor such as Inkscape The goal is to capture the complexity of the environment, while not necessarily drawing it to scale 2

3 Odometry Testing: Performed drive-forward test: Varied distance: 1 m, 2 m Varied velocity: 0.1 m/s, 0.25 m/s, 0.5 m/s Surface: lenolium in Halligan Performed start-stop test: Move interval: 2 sec; Stop interval: 1 sec Distance: 2 m Velocity: 0.1 m/s Surface: lenolium in Halligan Studied odometry data to understand the coordinate system of the robot Characterized the error in odometry data versus actual motion: Surprisingly, the errors were greater at smaller speeds At small speeds, there was a consistent drift to the left (this effect disappeared at higher speeds) The stop-start test did not increase the error significantly Mean error at 0.1 m/s was about 10% Mean error at 0.25 m/s was about 8% Mean error at 0.5 m/s was about 1% Implementation Plan: The initial implementation is going to work in the ADE simulator, using beacons as landmarks. This will allow us to implement and test the EKF separately from the landmark detection algorithm. The next step will be extracting landmarks from laser data. This can also be tested in simulation. Output useful mapping data. Test it in the simulator. Test everything in the real world. Tweak Kalman filter. Odometry data in the simulator: We managed to obtain odometry data from ADESim We modified ADESim to add Gaussian noise to the odometry data on each update. We may decide to change the noise model later High-level Design: Coded high level SLAM interfaces EKFServer - Provides best estimate of current location by performing EKF. LandmarkServer - Provides a list of currently seen landmarks and their location relative to the robot. MappingServer - Maintains persistent representation of map and outputs a map representation to the user. 3

4 Arch - Implements exploratory schema based behavioral code. Wrote config file to start all necessary servers. Learned to pass custom objects via java RMI. SLAM implementation: Landmark extraction EKF Fit lines to laser readings using RANSAC. Use estimated robot position from the EKF to find origin of global coordinate system. Project origin of coordinate system onto each fitted line, and take the projected point as a global point landmark. Report landmarks to the EKF only if they have been reobserved a sufficient number of times. Forget landmarks that have not been reobserved in a long time. Wrote a simple architecture to help find the distance between the robot s center of rotation and the position of the LRF. Added correction for the distance between the robot s center of rotation and the position of the LRF in simulation. The correction constant will be different on the real robot. Testing: offline visualization of RANSAC results, stepping through the iterations of the algorithm. Testing: online visualization of RANSAC results, written on top of LRFServerVis. Testing: in simulation, the robot consistently re-observes landmarks when driving and turning in a random manner. Prediction Update location from delta odometry. Use delta odometry to minimize odometry error (odometry most trusted for small deltas) - Done Update covariance with new robot prediction. Update robot location covariance and robot to feature covariance. - Done Correction Cycle through all old landmarks. Calculate Kalman Gain, predicted location, and delta location for each. Use each to update robot and landmark position. Update covariance with landmark position deltas and Kalman gains. Ignore old landmarks that have not been re-observed. Add new landmarks to state and covariance matrix. 4 Next Steps Finish EKF Implementation Andy Resolve inconsistencies in Slam for Dummies tutorial. Add covariance update to EKF correction step Test and tune EKF Implementation Andy 4

5 Create contingency plan if EKF fails to operate Both Figure out persistent mapping Constantin Create visualization of persistent mapping for the demo Constantin 5