Kalman Robot: Difference between revisions
Adamcolton (talk | contribs) No edit summary |
Adamcolton (talk | contribs) No edit summary |
||
| Line 12: | Line 12: | ||
The main board will be something like a RasPi. GPIO used to drive the two stepper motors. LiPo batteries. There will be a mating charge station. | The main board will be something like a RasPi. GPIO used to drive the two stepper motors. LiPo batteries. There will be a mating charge station. | ||
== Kalman Filter == | == Localization Software == | ||
The localizer software should be flexible so that I can publish it and other people can write software for it. | |||
I think the correct approach is going to be 3 levels. The outer takes position approximations from different sources and combines them. The second layer takes data from sensors and aggregates it into approximations. The third layer takes actual sensor readings and cleans it up for the models. | |||
=== Kalman Filter === | |||
The filter needs to be trainable so that it can learn about it's environment. | The filter needs to be trainable so that it can learn about it's environment. | ||
A good upgrade would be the ability to look for measurements that are no longer accurate and update them. | A good upgrade would be the ability to look for measurements that are no longer accurate and update them. | ||
=== Markov Localizer === | |||
Some of the data may work better with this. | |||
* https://vimeo.com/20528506 | |||
Revision as of 17:08, 4 October 2018
A robot that can be taught to navigate a finite space using a Kalman filter. The filter needs to position it in the X,Y axis as well as the X,Y rotation, so 3 dimensions.
Inputs
- Wheel Encoders
- GPS
- Accelerometer
- Wifi signal strength
- Sonar module
The design will be a tread based robot. The body can be 3D printed. The case will need to be water proof.
The main board will be something like a RasPi. GPIO used to drive the two stepper motors. LiPo batteries. There will be a mating charge station.
Localization Software
The localizer software should be flexible so that I can publish it and other people can write software for it.
I think the correct approach is going to be 3 levels. The outer takes position approximations from different sources and combines them. The second layer takes data from sensors and aggregates it into approximations. The third layer takes actual sensor readings and cleans it up for the models.
Kalman Filter
The filter needs to be trainable so that it can learn about it's environment.
A good upgrade would be the ability to look for measurements that are no longer accurate and update them.
Markov Localizer
Some of the data may work better with this.