Access to different cost values

[amirrazmjoo]

Hello there, I am new to crocoddyl and pinocchio, and I am trying to learn these packages by controling different aspects of a robot (panda in this case). So far, I've become familiar with controlling a frame position (gripper_pose) using crocoddyl and it works fast and fine, however, when I try to add cost about the position of center of mass, it does not affect the previous trajectory (i.e, the trajectory is as same as when there was no CostModelCoMPosition). To debug this and see what is going underline the code, I need to see how much the values of different costs are (e.g., gripper_pose cost, state_costs, ComTrack cost, etc.), but unfortunately I cannot find out how to do so. I would be so grateful if you could help me through.

[proyan]

Use

cost_val = solver.problem.runningDatas[ i ].differential.costs.costs["gripper_pose_cost"].cost`

to get the value of gripper_pose_cost for the ith problem node in your solver.

Use

cost_weight = solver.problem.runningModels[ i ].differential.costs.costs["gripper_pose_cost"].weight

to get the value of the cost weight.

The total cost that is being optimized = cost_weight * cost_val

Does this answer your question?

[amirrazmjoo]

Thanks for your reply, Some how yes, I mean, I did want to find cost of all nodes, but I can do it in a for loop.

Any way, in my case, I had to make some changes to receive the cost value: cost_val = solver.problem.runningDatas[ i ].differential[a number, in my case between 0 and 3].costs.costs["cost_name"].cost Can you explain me what that number is? (the number that I have to provide after differential)

[cmastalli]

Any way, in my case, I had to make some changes to receive the cost value: cost_val = solver.problem.runningDatas[ i ].differential[a number, in my case between 0 and 3].costs.costs["cost_name"].cost

This doesn't make sense. Note that the differential returns all the data contained in the differential action model.

Let us know if this clarifies your doubts, if so, could you close this issue?

[proyan]

can you return the output of dir(solver.problem.runningDatas[ i ].differential)?

[amirrazmjoo]

yes, it returns: ['class', 'contains', 'delattr', 'delitem', 'dict', 'dir', 'doc', 'eq', 'format', 'ge', 'getattribute', 'getinitargs', 'getitem', 'getstate', 'gt', 'hash', 'init', 'instance_size', 'iter', 'le', 'len', 'lt', 'module', 'ne', 'new', 'reduce', 'reduce_ex', 'repr', 'safe_for_unpickling', 'setattr', 'setitem', 'setstate', 'sizeof', 'str', 'subclasshook', 'weakref', 'append', 'extend', 'tolist']

[gfadini]

Hi, I think I saw what @amrn9674 is referring to when trying out the RK4 integrator. In that case there is indeed a list of data to perform the necessary computations.

In [10]: ddp.problem.runningDatas[500].differential[3]
Out[10]: <crocoddyl.libcrocoddyl_pywrap.DifferentialActionDataFreeFwdDynamics at 0x7fe5ec10a820>

If this is the case, then note that with the Euler integration method you perform a simple integration step, while with the RK4 method you need to compute multiple intermediate steps to perform a single integration. The datas hence should refer to those intermediate calculations. I hope this helps

[amirrazmjoo]

@gfadini yes, it makes sense. Thanks for your answer.

[gfadini]

your're welcome @amrn9674

[proyan]

Yup, that's right @gfadini. Thanks. @amrn9674 Since you are looking at the cost values, in an RK4 integrator, you would have to do gripper_cost = (cost[0] + 2. * cost[1] + 2. * cost[2] + cost[3])/ 6. to get the final cost. cost[ i ] corresponds to the data inside differential [ i ]