MPC sometimes crashed

[pmj110119]

I'm using the MPC example examples/isaac_sim/mpc_example.py, but I keep getting stuck in weird poses and crashing.

When the target moves faster and farther, the probability of MPC collapse is high. On my own robot (the joint speed is relatively low), the probability of crash is very high. On the franka-robot provided by curobo, the probability is smaller, but there is a chance of it happening.

(crash at 12s of the video)

curobo_mpc.mp4

What may be the reason for this?

[pmj110119]

And MPC seems to be less stable than the results of Motion Generation, the robot cannot accurately stop at the target pose, and the probability of collision is a little high.

I am very curious, for dynamic collision avoidance, whether the following method can be used:

1. Firstly use Motion Generation to obtain a trajectory, while the robot moves along the trajectory, "esdf" is continuously updated and checks whether a collision occurs on the trajectory.
2. If a collision occurs, a new trajectory is planned by Motion Generation again and executed.

Compared with MPC, will this method be more stable? I would be grateful if I could get your advice!

[balakumar-s]

From the video, it looks like the gripper has some free joints that can cause unstable dynamics. Can you make these joints fixed? This should make mpc tracking better.