Hi, I try to lift the object using a 4 finger hand. I found an interesting phenomenon: the finger-tips will sink into the object at the beginning of the simulation. Later it will go normal. In my mind, it’s related with solver parameters, so I set the solimp = “0.99, 0.999, 0.0001” to make it has less impedance (but I leave solref as default here). The phenomenon still the same. The video is here: Also, when I try to decrease the mass of the finger link using “mass” in the xml file, the whole process tend to crush when the simulation starts. I uploaded the xml file that I used. May I ask the reasons and how to reduce the phenomenon and the instability? Thanks!
Also, unexpected slide happens during the simulation, even the contact forces lies in friction cone. More specifically, I set friction coefficient as default (equals to 1) and set condim = 3. The actual contact forces (output through mj_contactForce) have friction angle [ 19.7936 21.0462 35.0621 10.1835] deg. However, after several seconds, the fingers begin to slide along the object. Friction is quite important for dexterous manipulation, and I am quite curious about what happens here.
Re slip, this is due to the soft contact model. See this new section in the documentation: http://www.mujoco.org/book/index.html#Soft Initial penetration is also due to soft contacts, but in a different way. The solref parameter specifies a virtual spring-dampler which stabilizes constraints and cancels penetration. Make solref[0] smaller. The default is 0.02, which is the sping-damper time constant. The smaller this is the stiffer the system becomes. This will also make it harder to integrate, so you may need a smaller time step. Note that internally solref[0] is clamped from below to 2*timestep, to avoid instabilities. So you can actually set it to 0, and get the stiffest system MuJoCo will allow for a given timestep. You can also experiment with the Runge-Kutta integrator for such systems.