Hi,
we are currently trying to simulate a quadruped with each leg configured as a “five bar linkage” system. Here is an side-view illustration of one leg:
We have cloned the “Anymal task” from the omiisaacgymenvs repo, and are trying to get our own quadruped to walk. But we are encountering some problems:
The first problem: The paw “joint constraint” is not very good. When running the RL training, we see that the “joint constraint” is not always satisfied. The two shanks are then sometimes not connected in the paw joint. This usually happens for just a short period of time, and it seems like that a spring force is pulling the links together again. Note that this paw joint has been excluded from articulation, so that we can simulate the closed kinematic chain.
The second problem:
Some of the quadrupeds are “taking-off” after some time. We are not sure if this has something to do with the closed kinematic chain or our collision meshes. It seems like they experience some large force somehow.
Hi! When you remove a joint from the articulation it follows the basic joint path in the solver - which is less stable than articulations. Unfortunately it is still required for closed loop chains, so we need to live with that.
However, you can adjust your simulation to run better - Remember that as a general solver, the simulation is trying to optimize to minimize the error subject to the constraints. One way to help the simulation to perform correctly is to tune the other gains on the robot so they don’t weigh in too much for minimal deviations, letting the solver push harder on the paw joint. You can also convert that joint to a D6, and force a high stiffness value to it so that the solver will also give it more weight when solving.
If you get stuck on that I can give you some video guidance.
