Mobile Robot Control 2020 Group 4: Difference between revisions
| Line 59: | Line 59: | ||
===Mapping=== | ===Mapping=== | ||
====Feature Recognition ==== | |||
====SLAM==== | |||
The chosen algorithm for localizing PICO and updating the map with real-world information is [http://robots.stanford.edu/papers/Montemerlo03a.pdf '''fastSLAM2'''], with the actual implementation largely supported by online lectures and research by [https://www.youtube.com/channel/UCi1TC2fLRvgBQNe-T4dp8Eg Cyrill Stachniss]. | |||
The core idea of fastSLAM (especially in its comparison to normal Extended Kalman Filter (EKF) Particle Filter SLAM approaches) is the concept that the relationship between a robot's motion and sensors, and the outside world, is such that given a perfect model, landmarks will always be in the same position. | |||
In a more reduced form, using a particle filter to sample the robot's path, this relationship still holds, allowing one to model the outside world as a list of landmarks with a much more static nature than an ever growing hypothetical map of the outside world (which is what normal EKF filters do). Due to this, the measurement covariance element of a single landmark is nothing more than a 2x2 matrix, which, given that it has to be inverted and the most efficient inversion algorithm is O(n3), is a very big deal. | |||
The core idea of fastSLAM2 is to patch a hole in fastSLAM1 where information is lost. Specifically, in the part of the algorithm that is described as 'sampling the pose' - making a prediction, for a particle, of the robot state in the next measurement iteration - fastSLAM2 does not make use of that iteration's measurements. However, by nature of the above relationship, measurements and odometry data are almost equivalently tied to the robot's behaviour. As such, fastSLAM2 includes measurement data when it predicts a robot's state, which, as the paper linked above explains, dramatically improves performance of the algorithm. | |||
====Map Updating==== | |||
===Navigation=== | ===Navigation=== | ||
Revision as of 16:28, 3 June 2020
Group Members
| Name | Student Number | |
|---|---|---|
| M. Katzmann | 1396846 | m.katzmann@student.tue.nl |
| B. Kool | 1387391 | b.kool2@student.tue.nl |
| R.O.B. Stiemsma | 0852884 | r.o.b.stiemsma@student.tue.nl |
| A.S.H. Vinjarapu | 1502859 | a.s.h.vinjarapu@student.tue.nl |
| D. van Boven | 0780958 | d.v.boven@student.tue.nl |
| R. Konings | 1394819 | r.konings@student.tue.nl |
Escape Room Challenge
As algorithm a wall-follower is chosen. If done successfully, PICO will follow the wall and reaches the finish. This algorithm is being chosen because by researching Escape Room project of previous years, this algorithm scored well, based on time and whether the PICO finished at all.
This algorithm consists of the following parts:
Step 1
- The PICO moves forward until it reaches a wall.
Step 2
- The PICO aligns to the wall.
Step 3
- Then the PICO follows the wall until it reaches a corner.
Step 4
- If this corner is convex, align to the wall which is perpendicular to the previous wall and repeat Step 3.
- If this corner is concave, drive a little further until the PICO is aligned to the exit, then go straight.
In the meantime the PICO is also being checked whether it is not too close to the wall and well alligned.
With this wall follower algorithm, the PICO achieved second place, and reached the exit in 27 seconds.
Hospital Challenge
[Under construction]
For the Hospital Challenge, the design of the algorithm is divided into four parts: interaction, supervisory, mapping and navigation. Each part has its own functionality and needs to work with the other parts. This functionality is explained below:
Interaction
These are all low-level interactions consisting mostly of calls to the IO API (which represents communication with PICO).
Supervisory
This section covers the governing control logic of the proposed system. The idea is that the supervisor makes choices, tunes parameters, and controls flow-of-command with in remaining code depending on its mode, and changes modes dynamically.
Mapping
Feature Recognition
SLAM
The chosen algorithm for localizing PICO and updating the map with real-world information is fastSLAM2, with the actual implementation largely supported by online lectures and research by Cyrill Stachniss.
The core idea of fastSLAM (especially in its comparison to normal Extended Kalman Filter (EKF) Particle Filter SLAM approaches) is the concept that the relationship between a robot's motion and sensors, and the outside world, is such that given a perfect model, landmarks will always be in the same position. In a more reduced form, using a particle filter to sample the robot's path, this relationship still holds, allowing one to model the outside world as a list of landmarks with a much more static nature than an ever growing hypothetical map of the outside world (which is what normal EKF filters do). Due to this, the measurement covariance element of a single landmark is nothing more than a 2x2 matrix, which, given that it has to be inverted and the most efficient inversion algorithm is O(n3), is a very big deal.
The core idea of fastSLAM2 is to patch a hole in fastSLAM1 where information is lost. Specifically, in the part of the algorithm that is described as 'sampling the pose' - making a prediction, for a particle, of the robot state in the next measurement iteration - fastSLAM2 does not make use of that iteration's measurements. However, by nature of the above relationship, measurements and odometry data are almost equivalently tied to the robot's behaviour. As such, fastSLAM2 includes measurement data when it predicts a robot's state, which, as the paper linked above explains, dramatically improves performance of the algorithm.
Map Updating
This section covers the system functionalities responsible for (intelligent) navigation. All functions here have read access to the world model’s internal map and LRF envelope. The navigation can be split into two parts: pathfinding and path movement.
Path Finding
For pathfinding, A* is chosen. A* is a pathfinding algorithm based and built on Dijkstra's Algorithm. Dijkstra's Algorithm looks at every available cell all around the start point, until it reaches the goal. A* is smarter, and attaches more value to the cells towards the goal. Because of this, A* needs fewer iterations and is faster.
Figure 1: Difference between Dijkstra's algorithm (left figure) and A* (right figure).
Path Movement
Path movement translates the points of the trajectory of the A* algorithm to a vector of necessary x and y speed values for the PICO robot to follow the trajectory.
List of Meetings
| Date/Time | Roles | |
|---|---|---|
| Meeting 1 | 01-05-2020
11:00 |
Chairman: Bas
|
| Meeting 2 | 04-05-2020
11:00 |
- |
| Meeting 3 | 08-05-2020
11:00 |
Chairman: Max
|
| Meeting 4 | 11-05-2020
10:00 |
- |
| Meeting 5 | 12-05-2020
10:00 |
- |
| Meeting 6 | 15-05-2020
11:00 |
Chairman: Dominic
|
| Meeting 7 | 18-05-2020
11:00 |
- |
| Meeting 8 | 22-05-2020
11:00 |
Chairman: Roel
|
| Meeting 9 | 27-05-2020
15:00 |
- |
| Meeting 10 | 29-05-2020
11:00 |
Chairman: Rob
|
| Meeting 11 | 02-06-2020
11:00 |
- |
| Meeting 12 | 04-06-2020
11:00 |
Chairman: Ananth
|
Links
These will probably be invite-based but a backup link could be useful.
Gitlab: https://gitlab.tue.nl/MRC2020/group4
Overleaf design document: https://www.overleaf.com/project/5eabeda72b3e4100010f8b40