Indeterminant linear system detected while working near variable

[ujas mandavia]

Hello everyone,

I'm encountering this issue when optimizing wheel odometry poses using wheel odometry factor. I visualized variable nodes and also factors associated to that node as well but couldn't find any issue there but still getting this error.

Indeterminant linear system detected while working near variable
554 (Symbol: 554).

Thrown when a linear system is ill-posed.  The most common cause for this
error is having underconstrained variables.  Mathematically, the system is
underdetermined.  See the GTSAM Doxygen documentation at
http://borg.cc.gatech.edu/ on gtsam::IndeterminantLinearSystemException for
more information.

And also when I create a factor graph using same wheel odometry data along with IMU measurements and add IMU factors along with optimizing imu bias nodes, I get the same error for the imu bias node and not the pose node. I don't know why this is happening

One more thing to add, these error occurs when I add pose prior factor and without the pose prior factor wheel odometry + imu optimizes perfectly.

Last thing, while add the pose prior factor, If I add x,y translation value 2 or grater than 2 and also the yaw value to be greater than 2 the factor graph optimizes with the covariance of "covariance": [0.05, 0.05, 0.05, 0.05, 0.05, 0.05] for the pose prior factor but if I add x,y translation and yaw value below 2 the optimizer throws the same error as above.

Thank you in advance for the help,

Ujas

[Matías Mattamala]

Hi Ujas,

Your problem may be undetermined if you have less factors than variables/degrees of freedom in your factor graph. There may be variables that are not constrained by any factor, so double check that all the factors are defined correctly are connected to the right variables. For example, imu biases introduce 6 more variables to optimize, so you need to ensure that you constrain them appropriately, for example, adding priors on the biases.

For the last question, what are the values that you are changing? the actual measurement/value for the prior or the prior's covariance? If the latter, larger covariances can make the factors vanish, so that could be the cause of the exception. If the former, I'm not sure why this could happen.

Cheers