Experiment catalog

The 16-case block force-limit experiment is complete: 0.2/0.4 N fail retention, 0.8/10 N hold and release, including ±2 mm offsets. Numerical validity and task outcome are separate; no hardware-fidelity claim.

Keep existing code paths to preserve asset references, reproduction commands and historical evidence. Each experiment connects a question, configuration, raw record, independent score, report and failures.

Experiment	Code	Question and status
RSS wiping and assisted retention	demos/rss-wiping/	The original uses compliant latch constraints to carry the sponge with the hand; matched no-latch runs measure contact-only holding, slipping and cleaning
Sequential cutting and knife-driven pushing	demos/tofu-knife-push/	Three tofu pieces and four banana pieces; explain separation through recorded knife-contact forces, using synthetic materials
Cross scoring	demos/tofu-cut/	Two shallow cuts and a connected base; distinguish elastic rebound from native cut history
Synthetic tofu halving	demos/tofu-cut/	Single-cut CPU TLMPM, dynamic blade reaction and separation after withdrawal; independent BladeMaster reference
Apple stem grasp	demos/apple-stem-grasp/	OpenArm + Wuji SDF-SDF frictional holding; distinguish default cases and historical perturbations
Robot cloth grasp	demos/cloth-folding/	Pinch, lift and release; historical intersections retained; one repaired 9 s development case passes, robustness untested
Contact and drives	demos/contact-benchmark/	Sliding, loading/unloading and transients; synthetic response is not hardware calibration
Basic cloth	Cloth engine comparison	Stretch, drape, collision and falling folds; separate material/solver capabilities
PhysX	demos/physx-contact/	Native rigid bodies, SDFs, joints and surface cloth; one success does not establish held-out performance
Genesis	#42	Scoped rigid evidence in #42; PBD cloth diagnostics retain failed momentum/grasp checks; tested frictional grasp profile rejected; equal-time response, connected folds and two held-out offsets recorded

src/dexlab/ contains shared task registration, scoring and recording. scripts/ contains setup, rescoring, documentation and research tools. Share code only for real duplication or invariants; this is not a new general simulation framework.

Apple control and dynamics

The right thumb/index pinch the stem while the left arm remains parked. Apple and stem form a 0.2 kg free rigid body. Known poses, IK and fixed joint targets drive the robot; the engine integrates actual object motion without attachment or object-position actuation. The continuous hold window is 11–14 s, with fruit contact allowed during approach.

The display camera follows recorded apple poses for playback only. There is no validated visual controller, stem fracture or hardware accuracy claim.

Implementation and parameters · Full inventory

Continuous close-up records

These are development-scene replays at the original pinned versions, not held-out success rates. The cloth recording preserves the detected table-intersection failure.

MuJoCo

SuperDex FP64

PhysX

Cloth: geometric failure

Hardware acquisition preparation

The protocol, empty template and log validator cover two UR7e arms and two identical CTAG2F90D grippers. Commands, measured feedback and independent references remain separate; unknown data stays unknown. The tool only reads files and does not control hardware. Format validity does not establish calibration.

Friction response development

The 16-run planar report retains all paired MuJoCo 3.14.0 / SuperDex FP64 outcomes, measured initial states, speed-bin force ratios and costs. Every run passes existing plane checks, while low-speed resistance differs. This is not matched material calibration or an engine-accuracy ranking.

Repaired cloth development case

One 9 s candidate passes the fixed v3 criteria; self penetration remains close to its limit. Metrics, failed controls, video provenance and exact command. This does not replace the historical failure above or establish robustness.

Initial Drake incline qualification

Official 1.57.0 / SAP / kLagged / hydroelastic: all six static and frictionless cases pass, three sliding cases fail, and the no-floor negative is rejected. The original coplanar failure is retained; 1 µm clearance belongs to a separate protocol. Reliable coverage under the new protocol remains unqualified. 报告 / Report.

Full experiment and historical index

Banana single-cut force: 20 samplings and 64× particles

CPU TLMPM / independent paper-value protocol

Is material resistance stable across sampling, timestep, contact scale and particle refinement?

Twenty baseline runs qualify numerically; the 64× case misses support contact within 0.5 s and gives 2.051 RMSE, so convergence and real-force accuracy remain unsupported.

Final cut directly into a side push: tofu and banana

CPU TLMPM / measured knife contact

Does visible blade-face contact drive the final piece apart?

Three tofu pieces and four banana pieces; the final cut transitions directly to a face push, displaying contact force and motion. Tofu adds Newtonian viscosity with retained elastic, half-step and no-push controls.

Historical: Synthetic tofu halving

CPU TLMPM / BladeMaster v2 reference

Do the two cut pieces move independently after withdrawal?

Three full-cut cases yield 50% mass pieces; negative controls stay connected. Force convergence and real-tofu calibration are unproven.

Historical: Two tofu cuts: three-method view

CPU TLMPM / author reference videos

Can the left piece move independently after two cuts?

Two cuts form three pieces; native knife contact pushes the left piece, with measured force and displacement. Author clips use a final velocity input, so inputs differ.

Crossed shallow tofu cuts

CPU TLMPM / guided orthogonal scoring

Does the lower bridge remain connected after crossing incisions?

Default and half-step cases form two shallow cuts while retaining one native component; the uncut control is rejected. Synthetic material, not food calibration.

Historical: Three sequential banana cuts

CPU TLMPM / persistent cut history

Do subsequent cuts preserve earlier discontinuities?

Progressive separation and complete controls in the experiment record. Surface/particles share physical states; no real-banana calibration.

Native LIBERO: diagnosis and change

LIBERO / robosuite 1.4.0 / MuJoCo 2.3.7 · Newton / elliptic

How can an existing benchmark expose and validate a change?

Stale observations identified; timestep trials retain successes, failures and regressions.

Normal & transient response

MuJoCo · SuperDex · historical PhysX

How do parameters change loading, unloading and transients?

Static calibration works; dynamics and mass transfer need separate checks.

Six-engine incline & friction

MuJoCo · SuperDex · Genesis · Newton Physics · PhysX · Drake

Where do static, sliding and low-friction cases differ?

Fixed-profile outcomes and observation validity are separate; failures remain.

Impact & contact onset

MuJoCo · SuperDex

How do timestep, stiffness and impact phase affect transients?

Existing phase/stiffness sweeps; smaller steps do not automatically remove model differences.

Fixture pinch, lift & release

Genesis · MuJoCo · SuperDex (separate protocols)

Why does the grasp hold or slip?

Genesis 0.4 N and 0.8 N configurations have different retention outcomes.

Apple-stem grasp · SDF contact

MuJoCo · SuperDex · historical PhysX

Can a grasp maintain contact on complex geometry?

Native continuous replays and historical independent checks; cohorts stay separate.

Cloth mechanics & robot grasp

MuJoCo · SuperDex · Newton Physics · Genesis · historical PhysX

When do deformation, contact and retention fail?

Historical compliant-grasp frame; mechanics, drop and grasp retain separate protocols and failures.

Mass / size transfer

MuJoCo · SuperDex

Do fixed parameters transfer to nearby scenes?

30 historical transfer records, 0/30 joint passes; failures guide selection.

Response error & cost

MuJoCo · SuperDex

What does lower error cost?

27 records separate native computation, recording and scoring costs.

Native & framework paths

MuJoCo / MJWarp · Newton / Isaac Sim

Why can matching cores still produce different trajectories?

GPU field differences remain after CPU alignment; parameter-map thumbnail is not trajectory evidence.

Pending / incomplete: broader pushing, in-hand rotation, active folding and grasp transfer remain in their existing Issues. Existing knife-driven pushing has its own protocol; folded drop is not active folding.

What we already know

Six-engine incline: shared conditions, different responses

A 40 mm, 64 g cube; static, sliding and nominal-zero-friction conditions at three timesteps. Curves and replays retain the original passes, failures and invalid records.

Synchronized replay, full-rate curves, parameters and original evidence

Normal response: change parameters or the solver?

The same MuJoCo fixture moves from about 80 kN/m toward the synthetic 20 kN/m target. Inspect loaded-window indentation, then the complete unloading trace; success belongs to the stated checks and conditions.

Synchronized replay, full-rate curves, parameters and original evidence

Pinch: 0.4 N slips, 0.8 N holds

The same Genesis fixture and object, changing the finger-joint force limit. Compare complete closing, lifting, holding and release, with an open-gripper negative control retained.

Synchronized replay, full-rate curves, parameters and original evidence

LIBERO: native task, timestep control and observation fix

Same actions, initial state, Panda and native success; original 2 ms versus unqualified 1 ms control. Heldout task success stays 10/10, screens change from 5/10 to 3/10, and cost doubles. Policy effects of observation refresh remain unverified.

Synchronized replay, full-rate curves, parameters and original evidence

Which engines and solvers are compared?

Engine	Roles in registered profiles (not peer algorithms)	Experiment versions and paths
MuJoCo	Constraint method: CG (nonlinear) · Newton · PGS Friction cone: elliptic · pyramidal Integration: Euler · implicitfast Unresolved role/identity: SolverMuJoCo; algorithm unqualified	3.11.0, 3.14.0, 3.15.0; native / historical paths in matrix
SuperDex	Deformable implementation: experimental-shell Outer nonlinear: BFGS · Newton · SR1 Inner linear: ASYNC_CG · AUGMENTED_CG · CG · CUDA_CG · CUDA_GMRES · EXPERIMENTAL_CUDA_SPARSE_CHOLESKY · EXPERIMENTAL_CUDA_SPARSE_LDLT · EXPERIMENTAL_CUDA_SPARSE_LU · GMRES · LDLT · LU · MINRES · PARALLEL_CG Linear selection: AUTO Friction regularization/model: C1_REGULARIZED · CINF_REGULARIZED Precision in label: FP64	1.0.0; native / historical paths in matrix
Genesis	Deformable implementation: PBD Constraint method: CG (nonlinear) · Newton Friction cone: elliptic · pyramidal Contact model: Signorini · convex Coupling path: rigid coupling Integration: approximate_implicitfast	1.4.3; native / historical paths in matrix
Newton Physics	Implementation/system: Featherstone · Kamino · SemiImplicit · VBD · VBD compliant · VBD legacy · XPBD Deformable implementation: ImplicitMPM · Style3D Dynamics backend: DVI · PADMM Integration: Euler Update/budget options: 100 sweeps	1.6.1 / Warp 1.18.0, 1.7.0.dev0 @ 2dee3234; native / historical paths in matrix
PhysX	Deformable implementation: surface cloth Constraint method: PGS · TGS Friction regularization/model: patch friction Update/budget options: external forces every iteration · friction every iteration Unresolved role/identity: historical readback incomplete (#126) · pending qualification	5.9.0; native / historical paths in matrix
Drake	Constraint method: SAP Contact model: hydroelastic · point Discrete contact approximation: kLagged · kSap · kSimilar Unresolved role/identity: not qualified	1.57.0; native / historical paths in matrix

Native and framework MJWarp are separate paths with their actual core identified. Newton Physics is an engine project; Newton is also an algorithm name. Tested SuperDex BFGS/SR1 with assembly period=1 execute Newton-equivalent steps; configuration names are not independent algorithm counts.

This is a research inventory including failures and unqualified entries. Execution, CPU/GPU, precision, full profiles and unsupported reasons: 完整矩阵 / Full matrix。

Roles interpret recorded names within each engine, not identical mathematical problems or evidence that every combination ran. AUTO is a selection policy; missing identities are not filled from defaults.

Comparison boundary: distinguish system-profile selection from controlled algorithm attribution. Read record integrity, numerical consistency, physical references and task outcomes separately; original scores and failure denominators are unchanged. Roles and evidence guide。

Thin-feature SDF geometry/response qualification remains #178; cloth material, collision/coupling and grasp baselines remain #179. Existing development passes are not calibrated cross-engine material comparisons.

MJWarp path comparison (MuJoCo core)

Runtime path	Actual physics core	Solver / contact path	Task acceptance
Isaac Sim6.1.0 local tag build / Newton1.5.0 / vendor Warp1.16.0	MuJoCo 3.11.0	MJWarp Newton / pyramidal / Newton contacts	not-run
Isaac Sim6.1.0 local tag build / Newton1.5.0 / vendor Warp1.16.0	MuJoCo 3.11.0	MJWarp Newton / pyramidal / MJWarp contacts	not-run
Native matched core / vendor Warp1.16.0; original and four-field aligned	MuJoCo 3.11.0	MJWarp Newton / pyramidal / MJWarp contacts	not-run

Model/clock diagnostics already exist; task acceptance retains the matrix state. Path differences #152 remain a separate research question.