Deadband incorrectly implemented in mixed pos/vel controller

[Hugo Elias]

This is a hard bug to explain, and I don't know what the fix is exactly.

When the joint is within the deadband, the output of the torque demand is zeroed. The problem with this is that the small position error is still fed into the position controller, which continues to integrate. This causes a hidden large velocity demand, and consequently a large hidden torque demand. As soon as the joint exits the deadband (for example due to coupling with another joint) the hidden torque demand is exposed, causing the joint to make a sudden large movement.

Toni and Hugo tried implementing some fixes including zeroing the position error (instead of the torque demand) when inside the deadband. This was an improvement, but still needs some work.