add benchmark for MyoSim MyoArm

[thowell]

add a benchmark for the MyoSim MyoArm musculoskeletal environment

a utility script benchmark/myosim/get_myosim.py is provided to clone MyoSim (pinned commit) from GitHub.

python get_myosim.py

---

testspeed

mjwarp-testspeed benchmark/myo_sim/arm/myoarm.xml --nworld=8192 --nconmax=16 --njmax=48 -o "opt.ccd_iterations=50" --event_trace=True --measure_alloc=True --memory=True

note: the default ccd_iterations (35) is too low and produces warnings

Loading model from: benchmark/myo_sim/arm/myoarm.xml...
  nbody: 40 nv: 38 ngeom: 161 nu: 63 is_sparse: True
  broadphase: NXN broadphase_filter: PLANE|SPHERE|OBB
  solver: NEWTON cone: PYRAMIDAL iterations: 100 iterative linesearch iterations: 50
  integrator: EULER graph_conditional: True
Data
  nworld: 8192 naconmax: 131072 njmax: 48

Rolling out 1000 steps at dt = 0.002...

Summary for 8192 parallel rollouts

Total JIT time: 0.59 s
Total simulation time: 11.17 s
Total steps per second: 733,257
Total realtime factor: 1,466.51 x
Total time per step: 1363.78 ns
Total converged worlds: 8192 / 8192

Event trace:

step: 1360.71
  forward: 1233.95
    fwd_position: 688.99
      kinematics: 80.03
      com_pos: 14.73
      camlight: 1.98
      flex: 0.17
      crb: 25.33
      tendon_armature: 9.97
      collision: 13.13
        nxn_broadphase: 4.10
        convex_narrowphase: 6.89
        primitive_narrowphase: 1.23
      make_constraint: 33.30
      transmission: 61.26
    sensor_pos: 0.17
    fwd_velocity: 145.89
      com_vel: 17.44
      passive: 15.31
      rne: 19.94
      tendon_bias: 27.18
    sensor_vel: 0.18
    fwd_actuation: 20.37
    fwd_acceleration: 128.34
      xfrc_accumulate: 4.74
    solve: 247.90
      mul_m: 13.04
    sensor_acc: 0.17
  euler: 126.21

ncon alloc:

mean     std      min    max  
------------------------------
16129.2  9024.78   8192  32768
2218.25  3018.07      2   8192
8850.36  1705.28   7770  15339
16338.3  164.723  15394  16384
  16384        0  16384  16384
13960.6  2936.51   8403  16384
8198.17  25.5985   8192   8367
   8192        0   8192   8192
   8192        0   8192   8192
   8192        0   8192   8192

nefc alloc:

mean   std       min  max
-------------------------
19.98   3.84182   16   32
 25.6  0.969536   23   27
25.81  0.868274   24   27
31.81    1.9733   26   33
30.69    1.0266   30   33
   30         0   30   30
30.02      0.14   30   31
   31         0   31   31
31.96   1.70833   31   35
   35         0   35   35

Total memory: 881.43 MB / 48640.12 MB (1.81%)
Model memory (2.01%):
 (no field >= 1% of utilized memory)
Data memory (97.99%):
 xmat: 11.25 MB (1.28%)
 ximat: 11.25 MB (1.28%)
 geom_xpos: 15.09 MB (1.71%)
 geom_xmat: 45.28 MB (5.14%)
 site_xpos: 47.53 MB (5.39%)
 site_xmat: 142.59 MB (16.18%)
 cinert: 12.50 MB (1.42%)
 ten_J: 79.56 MB (9.03%)
 wrap_obj: 29.00 MB (3.29%)
 wrap_xpos: 87.00 MB (9.87%)
 actuator_moment: 74.81 MB (8.49%)
 crb: 12.50 MB (1.42%)
 qM: 18.16 MB (2.06%)
 qLD: 18.16 MB (2.06%)
 efc.J: 72.00 MB (8.17%)

---

airspeed velocity

asv run myo_sim^! -b MyoArm

· Creating environments
· Discovering benchmarks
· Running 1 total benchmarks (1 commits * 1 environments * 1 benchmarks)
[ 0.00%] · For mujoco-warp commit 74371556 <main>:
[ 0.00%] ·· Benchmarking virtualenv-py3.13
[100.00%] ··· Setting up benchmark:144                                                                                                                                                                                                ok
[100.00%] ··· benchmark.MyoArm.track_metric                                                                                                                                                                                           ok
[100.00%] ··· =============================================== =====================
                                  function                                         
              ----------------------------------------------- ---------------------
                                jit_duration                    83.98443568497896  
                             solver_niter_mean                   1.3695615234375   
                              solver_niter_p95                         2.0         
                          device_memory_allocated                   941621248      
                                    step                        1363.148438744247  
                                step.forward                    1235.9470552764833 
                         step.forward.fwd_position              690.1426058029756  
                    step.forward.fwd_position.kinematics        80.12486679217545  
                     step.forward.fwd_position.com_pos          14.763125168428815 
                     step.forward.fwd_position.camlight         1.9792499891764237 
                       step.forward.fwd_position.crb            25.370499963173643 
                 step.forward.fwd_position.tendon_armature      9.984999742300715  
                    step.forward.fwd_position.collision         12.493625084971427 
                 step.forward.fwd_position.make_constraint      33.38687501491222  
                   step.forward.fwd_position.transmission        61.5145473384473  
                          step.forward.sensor_pos               0.1754999995569051 
                         step.forward.fwd_velocity              146.33549768768717 
                     step.forward.fwd_velocity.com_vel          17.502500037153368 
                     step.forward.fwd_velocity.passive          15.332375226535078 
                       step.forward.fwd_velocity.rne            20.019249903270975 
                   step.forward.fwd_velocity.tendon_bias        27.18824967450928  
                          step.forward.sensor_vel              0.17499999955816747 
                         step.forward.fwd_actuation             20.456375063076848 
                       step.forward.fwd_acceleration            128.8519998488482  
               step.forward.fwd_acceleration.xfrc_accumulate    4.746574093132949  
                          step.forward.sensor_acc               0.1781249995502776 
                             step.forward.solve                 247.7132497151615  
                                 step.euler                     126.66437595908064 
              =============================================== =====================

---

todo

• benchmark/myosim/get_myosim._clone is a direct fork of benchmark/kitchen/populate_scene._clone, make _clone a shared utility function?

[erikfrey]

Let's discuss how we want to design benchmarks going forward before we add this one. Might be time to make a few bigger changes.

[erikfrey]

Does this break running asv run main^! ?

Does the user have to manually run python get_myosim.py for the command above to work correctly?

[thowell]

yeah, probably need to run the script first

[erikfrey]

yes, in fact we should probably consider doing something like what playground does:

https://github.com/google-deepmind/mujoco_playground/blob/1bd244fa9c51fa7074c7a449934f556d898f7a15/mujoco_playground/_src/mjx_env.py#L92-L109

Perhaps it's time to support something like this so that we can keep these resources all outside of this repo.

[thowell]

created an issue #1001 for 'resource loader'