Help Needed: USD File Verification!

[Viru97]

Hey everyone!

I've assembled a robot USD file combining a UR10e arm with a Dexterous Inspire Hand, and I'm running into some frustrating physics instability issues when trying to train policy with rsl_rl framework.

Could anyone with experience in USD/robot physics/simulation please take a look at my file and help me verify if the joint definitions, masses, or collision geometries are set up correctly?
Any help would be greatly appreciated! 🙏 Let me know if you can assist! I will send files.

[e-ndorfin]

Hey - what sorts of physics instability issues are you running into? Do you mean when you're doing simulations there are some erratic / unexpected movements?

[Viru97]

Yes, either action rate explodes or joint velocity, due to that Mean value_function loss becomes inf.

something like this, i am using rsl_rl

################################################################################
                      Learning iteration 525/40500                      

                       Computation: 37935 steps/s (collection: 2.031s, learning 0.129s)
             Mean action noise std: 1.32
          Mean value_function loss: inf
               Mean surrogate loss: 0.0000
                 Mean entropy loss: 19.9778
                       Mean reward: 14.40
               Mean episode length: 275.13
    Episode_Reward/reaching_reward: 0.2150
    Episode_Reward/grasping_reward: 2.7344
Episode_Reward/object_lifting_reward: 0.0000
Episode_Reward/object_transporting_reward: 0.0000
Episode_Reward/object_placement_reward: 0.0000
Episode_Reward/final_success_reward: 0.0000
        Episode_Reward/action_rate: -603316068725154643968.0000
          Episode_Reward/joint_vel: -0.0381
    Metrics/command/position_error: 0.1785
 Metrics/command/orientation_error: 2.0644
     Metrics/command/command_stage: 0.0000
      Episode_Termination/time_out: 0.8772
Episode_Termination/object_out_of_reach: 0.1057
Episode_Termination/object_slipped: 0.0173
       Episode_Termination/success: 0.0000
Episode_Termination/abnormal_robot: 0.0000
  Episode_Termination/arm_collided: 0.0000
--------------------------------------------------------------------------------
                   Total timesteps: 2129920
                    Iteration time: 2.16s
                      Time elapsed: 00:00:58
                               ETA: 00:52:33

[e-ndorfin]

Have you done visualizations of the simulation? Does the arm successfully connect with the hand / are there some notable self-collision issues? Visualizing the environment would be useful.

You may try seeing effect of movement in the simulation via changing target position in the Physics Inspector or try teleoperation.

Then it will be easier to see why the action_rate and loss is spiking like this.

[Viru97]

Yes, i have connected arm, adapter and hand successfully.
I tested it using Gain-Tuner and Physics Inspector, everything seems working fine, but when there in collision between hand and the object then i get high action rate and joint velocity.