WIP: Use mc_rtc state observation pipeline and stabilizer #23
Conversation
| // FIXME: | ||
| // - resets does not respect the foot plan position when pausing walking (feet come back align | ||
| // dragging on the floor) | ||
| // - Do we still need the posture tricks? |
There was a problem hiding this comment.
Looks like the answer is yes, that's still how the controller implements Chest posture control (Section III.B). Without it, ankle compliance will comply uncompensated to the hip as we discussed last year stephane-caron#32, and that's extra work for the stabilizer (will likely fail the plank test then).
The indirect implementation via the anchor frame is not good, it's unnecessarily obscure. It is much better to do straight explicit feedback on the pelvis orientation like Kajita et al. proposed.
There was a problem hiding this comment.
Thanks for the detailed description of the role of pelvis control.
Here I was more referring to the fact that you are resetting the reference posture at every single iteration. I know you were doing this to prevent the FSM from resetting the posture to the current one on state transitions. For LIPMWalking itself this is not an issue, however many people have made custom controllers based on this one, and when used in other contexts such as manipulation, or when one explicitely wants to looks somewhere this becomes annoying. It also interferes with some FSM behaviours, such as the Interrupt button in the FSM (the robot comes back to halfsitting instead of keeping its current posture). Anyways I'll change the logic of this in a subsequent PR.
| // (6) updates that depend on realCom_ | ||
| netWrenchObs_.update(realRobot(), supportContact()); | ||
| stabilizer_.updateState(realCom_, realComd_, netWrenchObs_.wrench(), leftFootRatio_); | ||
| // XXX Default height is the same as that defined hidden in Stephan's |
There was a problem hiding this comment.
| // XXX Default height is the same as that defined hidden in Stephan's | |
| // XXX Default height is the same as that defined hidden in Stephane's |
😛
| stabilizer().run(); | ||
| pendulum().completeIPM(ctl.prevContact().p(), ctl.prevContact().normal()); | ||
| pendulum().resetCoMHeight(ctl.plan.comHeight(), ctl.prevContact().p(), ctl.prevContact().normal()); | ||
| controller().stabilizer()->target(pendulum().com(), pendulum().comd(), pendulum().comdd(), pendulum().zmp()); |
arntanguy
force-pushed
the
topic/stabilizer
branch
from
a4b8c2d
to
0f8af49
Compare
Based on the tuning done with Mehdi on branch topic/tune_hrp4j@6e62d56
Since yaml-cpp does not support, nor intend to support YAML merge anchors, this manually adds this ability to inherit from another robot's plans configuration to avoid duplication See jbeder/yaml-cpp#300
|
Closing in favor of #38 |
This PR removes the state observation and stabilization related code in favor of those integrated in
mc_rtc:TiltEstimatorsoon.So far this has been successfully tested in simulation, including walking to random targets and stair climbing. Still some further cleanup required, and tests on the real robots. No review required at this point.