On Tue, Sep 14, 2010 at 2:57 AM, clarkwu wrote: > Thanks for your prompt reply. > I also feel localization is not working well. > > As you suggested, I tested AMCL with just a joystick, and observed in rviz > that the localization is not reasonable, i.e., the laser scan deviate a lot > from map. hi XiaoJun, >From your video, it's hard to tell exactly what's going wrong, but I have two guesses: (1) There's a problem with your robot's odometry data. It almost looks like rotations are being reported backward. To get an idea for the quality of your odometry using rviz: * turn off the map and don't run amcl * set the fixed frame to 'odom' (or whatever your odometry frame is) * increase the decay time on the laser data (e.g., to 60 seconds) * using a joystick, slowly rotate the robot in place If your odometry data is good, you should see a crude "map" be constructed from the accumulated laser scans. Some amount of error is expected, but the scans should be visibly coherent. If they're not, then you should look at how you're computing the odom->base_link (or equivalent) transform. (2) amcl is not running fast enough to keep up with the incoming data. The particle cloud is updating infrequently and late with respect to the motion observed in the laser scan. What's the load on the machine running amcl? It's hard to quantify amcl's exact computational requirements, but it won't behave well if your CPU is 100% busy (many other things won't work well either). brian.