Engines, solvers and modeling assumptions#
Matching visual meshes, parameter names or joint targets does not establish matching physical problems.
Historical apple configurations#
Item |
MuJoCo |
SuperDex FP64 |
PhysX |
|---|---|---|---|
Contact path |
SDF contact search and soft constraints |
Surface-sampling integration and smooth penalty energy |
Native SDF contacts and TGS |
Physics step |
0.5 ms |
2 ms |
1 ms |
Key approximations |
SDF discretization and contact-point search |
Surface density and smooth penalty response |
SDF resolution, discrete contacts and torsional radius |
Evidence scope |
Default case and historical ten-scene cohort separate |
Default case and historical ten-scene cohort separate |
One development grasp, outside that held-out cohort |
Record collision geometry, contact law, integrator, constraint solver and drives separately. SDF defines geometry, not friction behavior or penetration accuracy. A smaller timestep need not improve every metric monotonically.
Parameter provenance and implementation · Contact/material research
Latest-stable qualification#
Before each new batch, qualify the latest stable native path or an officially compatible framework combination. Record core, binding and framework versions, hashes, precision and compute path separately; freeze throughout the batch. An older bundled core is not itself a blocker. Attribution experiments additionally match core versions; unmatched versions cannot isolate a framework effect.
2026-10-14: MuJoCo 3.16.0’s official environment and native compatibility are verified. The complete 4,000-step passive cloth trace first shows sampled midsurface intrusion at 22.5 ms (0.049681 mm); this is not material or grasp qualification. Complete task admission is tracked separately. The new-version qualification record and #202 retain failures and remaining acceptance. Historical data is not relabelled, and LIBERO’s officially compatible 2.3.7 path remains separate.
Release leads checked on 2026-10-04: MuJoCo 3.14.0, Genesis 1.4.3, Newton 1.6.0. This dated inventory alone does not establish task qualification. MuJoCo 3.14.0 now has the scoped cloth and friction development results in the results report; newer native incline cohorts are listed below and remain separate from those historical tasks.
The qualified Newton 1.6.1 package declares mujoco~=3.12.0 and mujoco-warp~=3.12.0 for its sim extra; ovphysx 0.6.3 is classified Alpha. Stable bindings, embedded cores and standalone SDKs need separate checks. Version task #41
Genesis scope#
Qualify rigid support, sliding and pinch/release first, then PBD thin cloth and coupling separately. MPM/FEM volume materials are distinct. Genesis’s Newton method is not Newton Physics. Unsupported SDF-SDF must not become an undisclosed convex surrogate. Qualification #42
GPU contact development evidence#
Six-device parameter checks on MuJoCo Warp 3.14.0 produced four passing and two failing configurations. The fixed 5 mm intrusion bound was unchanged; smaller timesteps did not monotonically improve penetration. This is a synthetic sphere/plane diagnostic, not apple or cloth qualification. Full results, plots and reproducible records.
Synthetic tactile observation#
Box/fixed-slab geometric occupancy is supported on the MuJoCo fixture; shear memory, optical rendering and hardware calibration are not. See all three batches and failed controls in the report.
Two-track admission and Drake results#
Drake 1.57.0 now has six incline configurations across three SAP approximations and point/hydroelastic contact: the reused cohort passes6/9, and new profiles pass0/9,5/9,0/9,3/9,2/9. Five new positives exceed the original momentum check; all failures remain. Source equations reconstruct old Lagged sliding forces within4e-13 N. Reliable task coverage is unchanged. Protocol, results and reproduction · Residual follow-up #165.
Native incline cohorts#
The coverage matrix retains the initial cohorts separately. Genesis 1.4.3 records five native CPU configurations. Newton 1.6.1 / Warp 1.18.0 records seven: XPBD and Kamino PADMM pass 1/9 positives each, DVI passes 4/9, and the remaining profiles pass none. The 49 invalid positives, two invalid VBD negatives and interrupted DVI attempt remain visible. An explicit continuation completes only the missing negative. These fixed configurations do not establish reliable coverage-v1 credit, equal tuning or an engine ranking.
Four native PhysX SDK 5.9.0 CPU patch-friction profiles now have a new incline cohort: 7/9, 7/9, 3/9 and 3/9 passes. All 13 numerically invalid positives remain, and all four negatives are valid and rejected. Protocol and evidence. Compatible framework cores and conversion remain separately scoped to #152; native results cannot be attributed to Isaac or historical cores.
Matched MuJoCo/MJWarp3.11.0 now runs natively beside Isaac Sim. The native support case also fails the frozen momentum check; aligning four GPU fields preserves that failure and still differs from the framework. Three independent aligned runs match exactly. CPU contact readback now rejects out-of-bounds addresses. Reliable task coverage is unchanged. Results and attribution limits.