## Re: [Playerstage-users] AMCL, problem with clustering in west direction

 Re: [Playerstage-users] AMCL, problem with clustering in west direction From: bots - 2011-01-05 17:27:15 ```Le mardi 04 janvier 2011 à 10:47 -0500, Rich Mattes a écrit : > > -----Original Message----- > > From: bots [mailto:bots.perroud@...] > > Sent: Sunday, January 02, 2011 6:26 PM > > To: playerstage-users@... > > Subject: [Playerstage-users] AMCL, problem with clustering in west > > direction > > > > Hi, > > > > I have been looking at the AMCL driver for sonar sensor. Going deeper > > in > > the code and testing it, I have found that hypothesis building has a > > negative bias for west direction. > > Angular positioning of the robot is more critical than with sonar, but > > even with stage and laser ranger, I got the following when the robot is > > heading west in a lobby. > > > > (player 2.1.3, stage v2.1.1) > > > > Hypothesis Count: 2 > > 0 (weight 0.578131): [pose: -2.95938,0.916891,-3.12727] > > 1 (weight 0.421869): [pose: -2.96589,0.914803,3.12949] > > > > The two hypothesis should be the same but due to the clustering method, > > there is a discontinuity at -/+pi > > > > When there is uncertainty about the orientation you get 3 hypothesis > > instead of two > > Hypothesis Count: 3 > > 0 (weight 0.815439): [pose: -3.97893,-0.364146,-3.04182] > > 1 (weight 0.182105): [pose: -3.90566,-0.322484,-0.132439] > > 2 (weight 0.00245614): [pose: -3.94218,-0.3651,3.13576] > > > > > > Because the distribution for the angle around pi is cut in two, it > > affect, I guess, the decision of the algorithm, because mean and > > variance are strongly affected. > > > > I am working with player-2.0.5 and I have modified the keys used for > > clustering using cos/sin instead of yaw, and it seems to be working > > better. > > > > May be this issue was known before, can somebody confirm the problem > > and > > if my solution make sense ? > > > > Thanks > > > > Philippe > > > > Problems around the +/- pi discontinuity do tend to crop up quite a bit, it > doesn't surprise me that this is happening. Given the results you provide, > it does seem like the discontinuity being handled incorrectly could be > causing your problem. I don't think anyone has caught this before (at > least, I don't think anyone's told us about it.) Can you send along your > solution to the problem, either by attaching the patch or by submitting a > patch to the patch tracker? If it works well I'll apply it to the code in > SVN. > > Rich > > > > ------------------------------------------------------------------------------ > Learn how Oracle Real Application Clusters (RAC) One Node allows customers > to consolidate database storage, standardize their database environment, and, > should the need arise, upgrade to a full multi-node Oracle RAC database > without downtime or disruption > http://p.sf.net/sfu/oracle-sfdevnl > _______________________________________________ > Playerstage-users mailing list > Playerstage-users@... > https://lists.sourceforge.net/lists/listinfo/playerstage-users > Hi, Rich My solution is working well on the player 2.0.5 I have modified. Patching : it is a good idea but I have to learn howto do it and also move the improvement to the SVN first, I guess. So, I need some weeks before submitting a patch. Philippe ```

 [Playerstage-users] AMCL, problem with clustering in west direction From: bots - 2011-01-02 23:26:41 ```Hi, I have been looking at the AMCL driver for sonar sensor. Going deeper in the code and testing it, I have found that hypothesis building has a negative bias for west direction. Angular positioning of the robot is more critical than with sonar, but even with stage and laser ranger, I got the following when the robot is heading west in a lobby. (player 2.1.3, stage v2.1.1) Hypothesis Count: 2 0 (weight 0.578131): [pose: -2.95938,0.916891,-3.12727] 1 (weight 0.421869): [pose: -2.96589,0.914803,3.12949] The two hypothesis should be the same but due to the clustering method, there is a discontinuity at -/+pi When there is uncertainty about the orientation you get 3 hypothesis instead of two Hypothesis Count: 3 0 (weight 0.815439): [pose: -3.97893,-0.364146,-3.04182] 1 (weight 0.182105): [pose: -3.90566,-0.322484,-0.132439] 2 (weight 0.00245614): [pose: -3.94218,-0.3651,3.13576] Because the distribution for the angle around pi is cut in two, it affect, I guess, the decision of the algorithm, because mean and variance are strongly affected. I am working with player-2.0.5 and I have modified the keys used for clustering using cos/sin instead of yaw, and it seems to be working better. May be this issue was known before, can somebody confirm the problem and if my solution make sense ? Thanks Philippe ```
 Re: [Playerstage-users] AMCL, problem with clustering in west direction From: Rich Mattes - 2011-01-04 15:48:08 ```> -----Original Message----- > From: bots [mailto:bots.perroud@...] > Sent: Sunday, January 02, 2011 6:26 PM > To: playerstage-users@... > Subject: [Playerstage-users] AMCL, problem with clustering in west > direction > > Hi, > > I have been looking at the AMCL driver for sonar sensor. Going deeper > in > the code and testing it, I have found that hypothesis building has a > negative bias for west direction. > Angular positioning of the robot is more critical than with sonar, but > even with stage and laser ranger, I got the following when the robot is > heading west in a lobby. > > (player 2.1.3, stage v2.1.1) > > Hypothesis Count: 2 > 0 (weight 0.578131): [pose: -2.95938,0.916891,-3.12727] > 1 (weight 0.421869): [pose: -2.96589,0.914803,3.12949] > > The two hypothesis should be the same but due to the clustering method, > there is a discontinuity at -/+pi > > When there is uncertainty about the orientation you get 3 hypothesis > instead of two > Hypothesis Count: 3 > 0 (weight 0.815439): [pose: -3.97893,-0.364146,-3.04182] > 1 (weight 0.182105): [pose: -3.90566,-0.322484,-0.132439] > 2 (weight 0.00245614): [pose: -3.94218,-0.3651,3.13576] > > > Because the distribution for the angle around pi is cut in two, it > affect, I guess, the decision of the algorithm, because mean and > variance are strongly affected. > > I am working with player-2.0.5 and I have modified the keys used for > clustering using cos/sin instead of yaw, and it seems to be working > better. > > May be this issue was known before, can somebody confirm the problem > and > if my solution make sense ? > > Thanks > > Philippe > Problems around the +/- pi discontinuity do tend to crop up quite a bit, it doesn't surprise me that this is happening. Given the results you provide, it does seem like the discontinuity being handled incorrectly could be causing your problem. I don't think anyone has caught this before (at least, I don't think anyone's told us about it.) Can you send along your solution to the problem, either by attaching the patch or by submitting a patch to the patch tracker? If it works well I'll apply it to the code in SVN. Rich ```
 Re: [Playerstage-users] AMCL, problem with clustering in west direction From: bots - 2011-01-05 17:27:15 ```Le mardi 04 janvier 2011 à 10:47 -0500, Rich Mattes a écrit : > > -----Original Message----- > > From: bots [mailto:bots.perroud@...] > > Sent: Sunday, January 02, 2011 6:26 PM > > To: playerstage-users@... > > Subject: [Playerstage-users] AMCL, problem with clustering in west > > direction > > > > Hi, > > > > I have been looking at the AMCL driver for sonar sensor. Going deeper > > in > > the code and testing it, I have found that hypothesis building has a > > negative bias for west direction. > > Angular positioning of the robot is more critical than with sonar, but > > even with stage and laser ranger, I got the following when the robot is > > heading west in a lobby. > > > > (player 2.1.3, stage v2.1.1) > > > > Hypothesis Count: 2 > > 0 (weight 0.578131): [pose: -2.95938,0.916891,-3.12727] > > 1 (weight 0.421869): [pose: -2.96589,0.914803,3.12949] > > > > The two hypothesis should be the same but due to the clustering method, > > there is a discontinuity at -/+pi > > > > When there is uncertainty about the orientation you get 3 hypothesis > > instead of two > > Hypothesis Count: 3 > > 0 (weight 0.815439): [pose: -3.97893,-0.364146,-3.04182] > > 1 (weight 0.182105): [pose: -3.90566,-0.322484,-0.132439] > > 2 (weight 0.00245614): [pose: -3.94218,-0.3651,3.13576] > > > > > > Because the distribution for the angle around pi is cut in two, it > > affect, I guess, the decision of the algorithm, because mean and > > variance are strongly affected. > > > > I am working with player-2.0.5 and I have modified the keys used for > > clustering using cos/sin instead of yaw, and it seems to be working > > better. > > > > May be this issue was known before, can somebody confirm the problem > > and > > if my solution make sense ? > > > > Thanks > > > > Philippe > > > > Problems around the +/- pi discontinuity do tend to crop up quite a bit, it > doesn't surprise me that this is happening. Given the results you provide, > it does seem like the discontinuity being handled incorrectly could be > causing your problem. I don't think anyone has caught this before (at > least, I don't think anyone's told us about it.) Can you send along your > solution to the problem, either by attaching the patch or by submitting a > patch to the patch tracker? If it works well I'll apply it to the code in > SVN. > > Rich > > > > ------------------------------------------------------------------------------ > Learn how Oracle Real Application Clusters (RAC) One Node allows customers > to consolidate database storage, standardize their database environment, and, > should the need arise, upgrade to a full multi-node Oracle RAC database > without downtime or disruption > http://p.sf.net/sfu/oracle-sfdevnl > _______________________________________________ > Playerstage-users mailing list > Playerstage-users@... > https://lists.sourceforge.net/lists/listinfo/playerstage-users > Hi, Rich My solution is working well on the player 2.0.5 I have modified. Patching : it is a good idea but I have to learn howto do it and also move the improvement to the SVN first, I guess. So, I need some weeks before submitting a patch. Philippe ```