Previously shell mode required cellcount=1 along at least one axis. This CL
adds support for cellcount > 1 in all three axes by pinning interior grid nodes
to the parent body and reconstructing their positions from boundary nodes via
Transfinite Interpolation (TFI).
PiperOrigin-RevId: 924314800
Change-Id: I8c2438f4866dd4133feed65f535a1ab69f0c9188
The field now stores either an approximate or exact diagonal of the constraint matrix A, so the name is made more general.
PiperOrigin-RevId: 924244954
Change-Id: I62b2f76531fb88b7b3bf96e6769940197596702b
Fixes#2313
This CL combines two complementary improvements to the Conjugate Gradient (CG) solver, significantly improving numerical precision, stability, and efficiency, particularly in single precision (float32).
1. Line Search Cost Evaluation Refactor: Previously, solver improvement was calculated by subtracting absolute costs: cost(alpha) - cost(0). In highly converged states or single precision, this is susceptible to catastrophic cancellation. We refactor PrimalSearch to compute the cost delta directly, dramatically improving precision.
2. Improved Solver Termination Condition: Near the float32 precision limit, line search deltas can occasionally be slightly negative due to numerical noise. Previously, any value below m->opt.tolerance (including negative values) triggered termination, halting the solver and locking in destabilizing steps. We update the termination condition to require positive improvement (0 < improvement < m->opt.tolerance), allowing the solver to continue iterating and recover stability.
Together, these changes yield substantial improvements, see reduced tolerances herein.
PiperOrigin-RevId: 924229669
Change-Id: Ic0bbefaed090f3a8b1e79ab8d45422c3e86fb56c
Two performance optimizations for interpolated flex objects:
1. Hoist loop-invariant stride calculations in `mju_cellLookup` out of nested loops.
This reduces multiplications from 16 to 6 (linear) and 54 to 12 (quadratic)
per vertex.
2. Extract and consolidate optimized 3D interpolation logic into a new
reusable utility `mju_evalBasisArray` in `engine_util_misc.c`. This function
uses nested loops and precomputed 1D shape functions to avoid expensive
dynamic `phi` calls and branching, and leverages stack-buffered outputs to
eliminate compiler pointer-aliasing barriers. We propagate this optimization
to both kinematics (`mju_interpolate3D`) and constraint setup
(`engine_core_constraint.c`).
Together these changes yield a ~50% overall speedup in `mj_fwdKinematics` for interpolated flexes with ~10k vertices in the collision meshes and ~10 nodes in the deformation grid.
PiperOrigin-RevId: 922198093
Change-Id: I9c804e65cd532e2f9ac6e9a71d01c30cd64186fa
The MSVC fallback path in safeAddToBufferSize() performed unchecked
arithmetic (type_size*nr*nc) on attacker-controlled values read from
.mjb binary model files. This could cause integer overflow, leading
to an undersized heap allocation followed by a heap buffer overflow
when data is copied into the buffer.
The fix adds manual overflow detection using SIZE_MAX/INTPTR_MAX
comparisons, matching the behavior of the existing __builtin_*_overflow
path used on GCC/Clang.
Also adds a regression test that crafts a binary model buffer with
overflow-inducing size fields and asserts safe rejection.
The normal component of the tactile sensor now reports the maximum penetration depth at each taxel, instead of a derived normal force. The depth is negated so that positive values indicate penetration.
PiperOrigin-RevId: 921980899
Change-Id: Ide4103c0aff465e25a81cfdda650f6a0e2da9e0b
mju_boxQP documents that only the lower triangle of the Hessian H is
read, but the gradient and search-direction updates inside
mju_boxQPoption still called the dense mju_mulMatVec, which reads the
upper triangle as well. This violated the documented contract and
prevented callers from safely leaving the upper triangle uninitialized.
Add a file-local mulMatVecSym helper that computes res = H*vec while
reading only the lower triangle of H (mirroring the convention of the
existing mulVecMatVecSym quadratic-form helper), and use it in place of
mju_mulMatVec in both call sites. Extend the BoxQP test suite with
UpperTrianglePoisoned, which fills the strict upper triangle of H with
NaN and verifies that the solver produces the same result as on the
clean symmetric input.
Reported by @lshdlut.
Fixes#3275
Introduce a fast path in mj_addFlexDerivatives for centered flexes where all nodes are attached to bodies with only 3 translational DOFs and no DOF-contributing ancestors. This allows for a direct scatter of K*vec, avoiding the overhead of sparse Jacobian construction and multiplication.
PiperOrigin-RevId: 916063477
Change-Id: I824728a5bdc3b975b04b46ecf22994922e307a8c
The global node indices stored in face_gidx are now checked to ensure they are within the valid range of node numbers, raising an error if any index is out of bounds.
PiperOrigin-RevId: 916025763
Change-Id: I1d2dcb11603137c337493541066b62b2ed2854c0
In `mj_flexPassive`, precompute and cache face positions, global indices, and corotational quaternions for all faces. This avoids redundant calls to `mju_flexGatherFaceState` for each bending edge, as each face can be part of multiple bending edges. This gives a 3-4x speedup in mj_passive.
PiperOrigin-RevId: 915912918
Change-Id: I3e1c00cf2d989f85398a0750a0faf359249cfe01
Skip flexedge velocity calculation for rigid and interpolated flexes, as they are not used in those cases. The computation is now done per-flex, only for the edges within each non-rigid, non-interpolated flex.
PiperOrigin-RevId: 915907396
Change-Id: I0beb1f98f0115c83da13fed184e89101f57563ab
This gives a 3x speedup in implicitfast.
Also cleanup old code that was used in the dense factorization of the stiffness matrix before we switched to CG.
PiperOrigin-RevId: 915900315
Change-Id: Id6973c4bfd7d371a43ec6db982703969b23a3550
The flex interp stiffness matrix is Negative Semi-Definite (NSD). When forming the RHS for implicit integration, the term involving the velocity and stiffness should be added, not subtracted. A new test is added to ensure energy stability for flex interp stretch stiffness with the implicitfast integrator.
PiperOrigin-RevId: 914845245
Change-Id: Iaaf0914909128e64e195f17cc5f2f344a8a43bc2
Standard flex (flex_interp=0) with thin-plate bending treated bending forces purely explicitly. This caused contact-induced vertex vibrations and non-physical energy injection for flat resting sheets, because the solver treated each vertex as an independent mass during contact and contact normals are orthogonal to stretch constraints.
Fix: extend the existing preconditioned CG solver to include the constant bending stiffness K_bend in the implicit operator via matrix-free mat-vec.
PiperOrigin-RevId: 914774020
Change-Id: I45e0d6749abb6f873566203bccae956514b2576b
with a preconditioned Conjugate Gradient (CG) solver that operates
on the full system matrix.
The previous approach extracted flex DOFs into a reduced banded system,
factored it separately, and overwrote the global solve. This required
precomputed bandwidth (makeFlexBandwidth), parent-joint detection,
coupling corrections, and a FlexInterpContext struct — and only worked
for standalone flex trees without parent joints.
The new CG solver uses the already-factored global system (M - h*qDeriv)
as a preconditioner and adds the flex stiffness contribution via
matrix-free products (mjd_flexInterp_mulKD/mulK). This handles any
kinematic configuration — including flexes attached to articulated
chains or with parent joints — without sparsity pattern restrictions.
Before (`bunny_multicell`):
```
Simulation time : 50.80 s
Steps per second : 197
Realtime factor : 0.20 x
Time per step : 5080.3 µs
CG iters / step : 3.16
Contacts / step : 31.04
Constraints / step : 124.15
Degrees of freedom : 178
Dynamic memory usage : 0.4% of 100M
```
After:
```
Simulation time : 9.52 s
Steps per second : 1051
Realtime factor : 1.05 x
Time per step : 951.7 µs
CG iters / step : 3.21
Contacts / step : 30.90
Constraints / step : 123.61
Degrees of freedom : 178
Dynamic memory usage : 0.3% of 100M
```
PiperOrigin-RevId: 913758038
Change-Id: If5aa617b2d535c86aec9bd71c9e0003a2b38bdd7
The coordinates for flex interpolation are now computed using the absolute values of the vertex weights. The sign of the first vertex weight is then applied to the resulting barycentric weights. This correctly handles cases where the flex is both the first and the second entity in the contact pair.
PiperOrigin-RevId: 913590928
Change-Id: I970b35fba3d209e13b5b33bb5f945e3c6a43d487
The normal jump residual in flex elasticity calculations now rotates the rest-frame normal jump into the current frame using the face's corotational quaternion before subtracting the jump from the current normal difference.
PiperOrigin-RevId: 913242127
Change-Id: Ia62b28ecccd59e79225737d8225ef4131b333c96
- Compute efc_J_rowsuper incrementally in mj_addConstraint instead of post-hoc via mju_superSparse.
- Better exploitation of supernodes in A matrix pipeline: precount and fill skip redundant chain traversals for supernode rows.
- Redundant B_rowsuper computation via mju_superSparse is eliminated (indentical to efc_J_rowsuper).
PiperOrigin-RevId: 910787825
Change-Id: I6eda9996659602b7051ee1090aeedb862603c84e
This change adds a new passive force computation for flexes with elastic2d="bend" and dof="trilinear". The bending energy is based on the squared difference of normals between adjacent face elements at their shared edge midpoint. The edge data is precomputed during model compilation and stored in flex_bending.
PiperOrigin-RevId: 910772638
Change-Id: I3b12c7b7f1ba6ac1875df495d89e8cfec921ca80
Kyle Bayes