Commit 7c4ad5a
committed
WIP H1/H2: Newton actuator stiffness bump diagnostic
H1 result (stiffness 50/damping 5): mean reward 617.8 at iter 200, vs Newton
baseline at stiffness 1.0/0.1 = 27.7, vs PhysX at stiffness 1.0 = 246.7.
Conclusion: the Newton baseline 10x reward gap was caused by under-tuned
actuator gains, NOT by the actuator-binding bug (Newton issue #2585).
Wiring is functional; the per-DOF joint_target_ke/kd/pos arrays IsaacLab
Newton writes to ARE consumed by MJWarp.
Why the discrepancy with PhysX at the same nominal stiffness=1.0:
- PhysX path layers fixed_tendons_props(limit_stiffness=30, damping=0.1)
on top of the implicit drive — adds tendon-coupled torsional stiffness
Newton's MjcTendon parsing does not contribute equivalently.
- PhysX runs solver_position_iteration_count=8 per substep, amplifying
effective stiffness; Newton's MJWarp implicit-PD step is single-pass.
Current value 20.0/2.0 (H2) is the mid-range probe — still WIP.1 parent ccac624 commit 7c4ad5a
1 file changed
Lines changed: 13 additions & 8 deletions
Lines changed: 13 additions & 8 deletions
| Original file line number | Diff line number | Diff line change | |
|---|---|---|---|
| |||
180 | 180 | | |
181 | 181 | | |
182 | 182 | | |
| 183 | + | |
| 184 | + | |
| 185 | + | |
| 186 | + | |
| 187 | + | |
183 | 188 | | |
184 | | - | |
185 | | - | |
186 | | - | |
187 | | - | |
| 189 | + | |
| 190 | + | |
| 191 | + | |
| 192 | + | |
188 | 193 | | |
189 | 194 | | |
190 | | - | |
191 | | - | |
192 | | - | |
193 | | - | |
| 195 | + | |
| 196 | + | |
| 197 | + | |
| 198 | + | |
194 | 199 | | |
195 | 200 | | |
196 | 201 | | |
| |||
0 commit comments