Yes, I have concluded that Bayesian Statistics is magic. You can flip around terms inside P(x...| y....) several times and you will magically get the form you desire. XD

Now I understand the idea and the algorithm of FastSLAM clearly. Great lecture.

Thank you Tre Cool for explaining SLAM so well

I wish I could see the laser pointer clearer in these videos :(

This prof is really good unfortunately I joined FReiburg after he left

First things first, thanks for this amazing Material, however , i have a question. It is about your Pseudo Code, specifically the Variable Q, and its calculation. H is, the Jackobian, so far so simple I understand most of what it is, but I dont entirely understand what the sigma sign in the Line 14 (Slide 30) creates the sum of. Qt is, as i understand, the noise i expect from the sensor, but what dimension would that have? the same dimensionality as the Sensor Input?

Why is at 48:00 the proposal distribution p(x1:t | z1:t-1, u1:t) and not the Motion Model as in the MCL or are those equivialent?

awesome video lecture , i try bailey code but it give me uncorrect result when i change particle numbers ! can you help me to run it correctly ?

Dear professor, I have a small clarification. Why can't we do the same type of decoupling for EKF. That is, why can't we assume landmarks are independent of each other in EKF and only do Mx2x2 K.F. updates?

@CyrillStachniss

8 жыл бұрын

Because it hold only given the poses, so they are only conditionally independent.

I dont understand how the Sigma covariance matrix in the 2x2 EKF for each landmark (28:09) is calculated. How is the variance matrix R of motion integrated to the Sigma matrix of this 2x2 EKF?

@hathuytu

5 жыл бұрын

it looks like the Sigma covariance matrix in FastSLAM is counted for the uncertainty of the measurement and the map only. The uncertainty of the robot pose is embedded in the sets of particles.

Hey i got a little question about the PseudoCode of the FastSlam. In line 5 it seems like we update only one landmark per partikel. Does that mean every timestep, we only observe one landmark? In your update example there are two landmarks updated? i would have made a for loop over all landmarks in Line 5, or is that wrong? Thanks for every help i get ;)

@CyrillStachniss

10 жыл бұрын

I guess you are referring to the function FastSLAM1.0_known_correspondence. This function FastSLAM1.0_known_correspondence is called for every landmark observation, one after the other. For example, if the robot sees 4 landmarks in its current laser scan, this function is called four times. I hope this answers the question. Best, Cyrill

@GCOMRacquet

10 жыл бұрын

Cyrill Stachniss Thanks for the answer. I guess for simplicity, this algorithm assumes only one landmark is measured in every laser scan? Otherwise i would have to sample over all Partikels and sample the Pose for every landmark in a timestep, that wouldn't be very efficent or do i get something wrong? Best, Sascha

@heltok

9 жыл бұрын

Stu dent Yeah, you are right. The explanation is on page 449 in Probabilistic Robitics, in the algo they assume only one observed new feature. A more general solution is on page 461 which has a for loop over all the observed features.

@gideonkock8426

3 жыл бұрын

@@CyrillStachniss and how does this work for the importance weight? In your example, do you use the max weight of the 4 landmarks you have for the particle in consideration?

@gideonkock8426

3 жыл бұрын

@@heltok the one on 461 is with unknown correspondencies and only updates the ML feature.