Complete solver-audit inventory#

All 50 public RSS entrypoints#

Forty entries have task-scoped assertions; ten are tutorials, rendering or hardware mocks. Goal wrappers and their scene may share a trajectory: these are not fifty independent manipulation types. A passing expected-failure control is not successful manipulation.

Entrypoint	Current outcome	Scope	Upstream test
00_getting_started/ 00_hello_world.py	Scoped assertions pass	Original scene, controller and assertions; no hardware calibration	Original test
00_getting_started/ 01_core_concepts.py	Scoped assertions pass	Original scene, controller and assertions; no hardware calibration	Original test
01_cloth/ 00_towel.py	Scoped assertions pass	MuJoCo/VBD one_way; cloth reactions are not applied to the arm	Original test
01_cloth/ 01_fabrics.py	Scoped assertions pass	Construction and 20 steps; no manipulation acceptance	Original test
01_cloth/ 02_engines.py	Scoped assertions pass	Default Newton construction and 20 steps; not every external engine	Original test
01_cloth/ 03_scrub_a_stain.py	Scoped assertions pass	one_way plus force readback only; stain removal is not calibrated wear	Original test
01_cloth/ 04_your_own_cloth.py	Scoped assertions pass	MuJoCo/VBD one_way; cloth reactions are not applied to the arm	Original test
02_soft/ 00_sponge.py	Scoped assertions pass	SuperDex soft body; lift/hold without latch, distinct from loaded wiping	Original test
02_soft/ 01_tet_mesh.py	Scoped assertions pass	Tetrahedral import; construction and 20 steps	Original test
02_soft/ 02_your_own_soft.py	Scoped assertions pass	Custom soft-body type; construction and 20 steps	Original test
03_materials/ 00_rope.py	Scoped assertions pass	VBD contact history and sticky contact; gripper friction 60	Original test
03_materials/ 01_rope_memory.py	Scoped assertions pass	Friction grasp and plastic memory; elastic negative control retained separately	Original test
03_materials/ 02_twist.py	Scoped assertions pass	Friction-driven twist; stiff-rope slipping is an expected control outcome	Original test
03_materials/ 03_knot.py	Scoped assertions pass	SuperDex with reaction feedback off; no-self-contact negative control retained	Original test
03_materials/ 04_band.py	Scoped assertions pass	SuperDex closed-loop rope, reaction feedback off; broken-band control retained	Original test
03_materials/ 05_paper.py	Scoped assertions pass	VBD one_way plus slide_feedback; prismatic-joint feedback gives no net arm reaction	Original test
03_materials/ 06_bag.py	Scoped assertions pass	SuperDex self-contacting shell; grasp does not pin nodes	Original test
03_materials/ 07_brush.py	Scoped assertions pass	Original brushing task and local cleaning model	Original test
03_materials/ 08_tape.py	Scoped assertions pass	30 g rigid pull tab attached to cloth; not direct thin-film grasping	Original test
03_materials/ 09_adhesion.py	Scoped assertions pass	Original adhesion and pressing task; no calibration to real adhesive	Original test
03_materials/ 10_cutting.py	Scoped assertions pass	Original cutting model; blunt-blade negative control retained	Original test
03_materials/ 11_granular.py	Scoped assertions pass	MPM grains; original rice-sweeping task	Original test
03_materials/ 12_liquid.py	Scoped assertions pass	MPM viscous liquid; grasp and tilt a cup	Original test
03_materials/ 13_dough.py	Scoped assertions pass	MPM plasticity; retained deformation after pressing	Original test
03_materials/ 14_water.py	Scoped assertions pass	Genesis SPH cup pouring; shallow-tilt and offset negative controls retained	Original test
04_extending/ 00_add_a_robot.py	Scoped assertions pass	Original scene, controller and assertions; no hardware calibration	Original test
04_extending/ 01_new_kind.py	Scoped assertions pass	Custom chain type; construction and 20 steps	Original test
04_extending/ 02_suction.py	Scoped assertions pass	Default suction weld; separate wrench load-threshold checks, not pure friction	Original test
05_studio/ 00_your_own_panel.py	Scoped assertions pass	Original scene, controller and assertions; no hardware calibration	Original test
05_studio/ 01_render.py	Scoped assertions pass	Still/video file validity, not visual control	Original test
06_tasks/ 00_your_first_task.py	Scoped assertions pass	Original scene, controller and assertions; no hardware calibration	Original test
06_tasks/ 01_goals_and_rewards.py	Scoped assertions pass	Goal evaluation tutorial, not a manipulation episode	Original test
06_tasks/ 02_episodes.py	Scoped assertions pass	Initial-state sampling/reset tutorial	Original test
06_tasks/ 03_granular_scored.py	Scoped assertions pass	Original task-goal assertions	Original test
06_tasks/ 04_brush_scored.py	Scoped assertions pass	Original task-goal assertions	Original test
06_tasks/ 05_suction_scored.py	Scoped assertions pass	Original task-goal assertions with suction retention	Original test
06_tasks/ 06_bag_scored.py	Scoped assertions pass	Original task-goal assertions	Original test
06_tasks/ 07_band_scored.py	Scoped assertions pass	Original task-goal assertions	Original test
06_tasks/ 08_tape_scored.py	Scoped assertions pass	Original task-goal assertions; rigid pull tab	Original test
06_tasks/ 09_cutting_scored.py	Scoped assertions pass	Original task-goal assertions	Original test
06_tasks/ 10_water_scored.py	Scoped assertions pass	Original task-goal assertions	Original test
07_policies/ 00_pick_and_place.py	Scoped assertions pass	Fixed GaP action graph; no training	Original test
07_policies/ 01_check_and_retry.py	Scoped assertions pass	Fixed GaP graph; at least 3 successes in 4 episodes	Original test
07_policies/ 02_without_privilege.py	Scoped assertions pass	Example camera perception graph; not a general visual policy	Original test
07_policies/ 03_many_worlds.py	Scoped assertions pass	Fixed graph; at least 14 successes in 16 worlds	Original test
08_demos/ fold_the_towel/ fold_the_towel.py	Scoped assertions pass	one_way; documented model/geometry differences from the paper reference	Original test
08_demos/ route_the_cable/ route_the_cable.py	Scoped assertions pass	Friction and contact history, gripper friction 60; port orientation differs from reference	Original test
08_demos/ wipe_the_bar/ wipe_the_bar.py	Scoped assertions pass	Latch-assisted retention after contact readiness; not pure-friction wiping	Original test
09_teleop/ 00_leader_arm.py	Scoped assertions pass	Mocked hardware and simulated follower grasp checks; no hardware motion	Original test
09_teleop/ 01_force_feedback.py	Scoped assertions pass	Mocked force-feedback interface; no hardware force-feedback acceptance	Original test

Paper showcases and public examples have different denominators#

The README splash description separately mentions 48 tasks, but no one-to-one executable manifest for that number was found in the inspected public checkout. The 50 entrypoints, 18 paper showcases and splash count are not interchangeable.

Paper v2, Figure 5 shows 18 tasks. The frozen public repository has 50 executable example entrypoints and 12 helpers; these are not the same set. No shipped .task/.suite task content or additional configured task roots were found. The four related public demonstrations below explicitly differ from the paper/reference scenes and cannot establish exact reproduction.

The project website’s old pipeline has drivers but depends on unavailable gallery recordings and side directories. Its README-linked robosimstudio/demo repository returns 404; this does not reveal whether it is private or removed. All 18 exact paper-scene identities remain unverified and blocked by unavailable scenes/original data, rather than classified as solver failures.

Paper showcase	Related accessible code
Tool sorting	Website syringe driver; original scene/data unavailable
Sharps disposal	Syringe driver describes tool_sort/syringe into a sharps box; paper identity unverified
Basket packing	Complete scene/acceptance entry not located
Gear meshing	Website gears driver; original scene/data unavailable
Tool hanging	Website scissors driver; original scene/data unavailable
Plate racking	Website plates driver; original scene/data unavailable
USB-C insertion	Website USB-C driver; original scene/data unavailable
Cable routing	Website weave driver; related public route-the-cable example
Drawer unlocking	Complete scene/acceptance entry not located
Conduit threading	Website conduit driver; original scene/data unavailable
Crate flipping	Website crate driver; original scene/data unavailable
Leaf pressing	Complete scene/acceptance entry not located
Towel folding	Website towel driver; related public fold-the-towel example
Phone film peeling	Website peel driver; public tape example uses a rigid tab
Carton tape cutting	Complete scene/acceptance entry not located
Sponge wiping	Website wipe driver; related public wipe-the-bar example
Strip polishing	Complete scene/acceptance entry not located
Heat transfer printing	Complete scene/acceptance entry not located

The website’s separate prerecorded interaction pipeline explicitly latches the film tab, maintains ParticleDragger pins for the towel, and pins the whip. Its README distinguishes that pipeline/site from the currently published www frontend. These source facts do not establish how every current published video was produced. Neither rendered appearance nor unavailable state provenance proves friction-only retention.

Sources: public demo differences, website README, reconstruction dependency, film pipeline.

The pipeline also preserves author-reported historical regressions. The conduit driver says the rebuilt scene changed from 54 to 59 bodies; two 30 September attempts failed acquisition or advanced only 21 mm. The weave driver says the old weave5 scene was replaced; the nearest port_medium variant completed three of four pegs and missed the port, reward 0.5. These are reported scene/controller regressions, not local reproductions or uniquely attributed solver defects. Sources: conduit, weave.

All existing evidence directories#

The following 30 directories were reviewed at DexLab c0c0aa6. They include successful controls, rejected records and incomplete qualification. This is a review of existing evidence, not a new execution of every historical trajectory.

Evidence directory	Assessment
cloth-studio	Old online grasp criteria passed despite table intrusion found by a separate geometry check; later settling development is a separate cohort.
cutting	Index and presentation of synthetic TLMPM references; not an additional six-engine comparison.
drake-contact-paths	Six contact paths. Lagged friction is explained by its previous elastic normal force; five small impulse discrepancies remain unresolved.
drake-incline	Historical kLagged/hydroelastic configuration passes 6/9; later path diagnostics explain part of its difference from ideal Coulomb friction.
elastic-impact	MuJoCo 3.15 passes 18/18 terminal checks; these do not check every intermediate penetration.
engine-qualification	Version/API admission; MuJoCo 3.16 cloth settling still exposes local table intrusion.
force-limit	Low actuator force fails retention; larger force holds and releases. Expected load failure, not a solver defect.
framework-mjwarp	Native and wrapper paths remain different after core/parameter alignment; unique cause unresolved.
genesis-cloth	Missing legacy PBD–rigid tangential friction, delayed reaction and failed cloth retention.
genesis-incline	Effective friction floor and numerical stopping effects are partly explained; failed configurations remain retained.
impact-discrete	54/54 stored trajectories match the reconstructed discrete equations; this does not establish continuous-time physical accuracy.
impact-phase	30/36 terminal passes but 14/36 joint passes; arrival phase matters. Nine records are reused history.
impact-stiffness	24/27 terminal passes but 2/27 joint passes; stiffness, step size and penetration interact. Nine records are reused.
incline-comparison	Matched geometry, different contact models; no calibrated-material accuracy ranking.
incline-diagnosis	Twelve MuJoCo impratio diagnostics do not satisfy every frozen requirement.
incline-friction	Nine of eighteen MuJoCo configurations pass; soft-contact drift and sliding error retained.
libero	Task replay success and contact consistency differ. A stale post-settling observation is a wrapper defect, not a core solver defect. The 130-task parent scope remains incomplete.
mujoco-incline	Six fixed native configurations and numerical rejections; no universal inability claim.
newton-incline	Seven CPU FP32 profiles; 49 positive numerical/representation rejections. The engineering guard is not a derived floating-point error bound.
newton-xpbd	Sphere/plane positive, repeat and disabled-contact controls pass.
physx-incline	PGS/TGS fixed profiles, with thirteen positive numerical rejections; small TGS residuals are not fully explained.
pinch-impulse	Tighter stopping reduces force/momentum residual while creep persists.
pinch-load	Eighteen MuJoCo records: three passes, nine physical failures and six numerical rejections.
rss-wiping	Three matched original/no-latch pairs demonstrate retention assistance dependence.
studio	Replays existing records; creates no independent physical result.
superdex-incline	Twelve configurations retain static creep; some CG steps stop without closure. BFGS/SR1 with assembly=1 do not constitute distinct effective methods here.
synthetic-tactile	Gravity compensation masks support in batch 1; endpoint failure in batch 2; bounded success after settling in batch 3. Control/observation semantics, not validated tactile sensing.
unisim-contact-migration	Twenty-one native/adapter rescoring pairs agree, including low-force failures.
unisim-reuse	Native grasp, open-hand controls and force bookkeeping support adapter reuse, not a new solver.
visual	Media/state provenance, not additional task-success trials.

All directories: frozen evidence tree.

All 15 demo directories#

Demo directory	Assessment
apple-stem-grasp	Historical complete 14 s MuJoCo/SuperDex scripted checks; not Wuji in-hand or real-apple qualification.
banana-cut	Synthetic three-cut/four-piece TLMPM reference, with original failures retained.
banana-force	Sampling and sensitivity records exist; 64× particle run does not reach support within 0.5 s and fails acceptance. Normalized force RMSE 2.05054 is not convergence.
cloth-benchmark	Seven configurations, 105 historical trials; nominal materials are not calibrated equivalents.
cloth-folding	Old table intrusion and bimanual failures retained; a later 9 s single-grasp development passes. Bimanual folding remains unqualified.
contact-benchmark	Thirty-eight development experiments, twenty-five passes and thirteen failures; includes edge discretization, normal response, transient cost and transfer diagnostics.
cut-push	Blade-driven pushing with no-push and material controls; synthetic reference, not real-food accuracy.
cutting-refinement	Mechanical checks for refined tofu/banana scenes; no general material convergence claim.
libero-contact	Native demonstration-action replay and contact diagnostics; no trained-policy or real-robot result.
physx-contact	Native primitive contact checks pass; missing UniSim contact reporting was repaired in the adapter. Separate cloth records include surface crossing and unavailable force APIs.
rss-wiping	Matched latch ablation; compliant node assistance is neither friction-only grasping nor a rigid weld.
tofu-comparison	Historical two-cut and one-sided velocity diagnostic; a different protocol from blade pushing.
tofu-cut	Single cut with uncut/partial-cut controls and dynamic reaction.
tofu-knife-push	Four pushing/control conditions with matched pre-push states.
tofu-scoring	Shallow crossing cuts preserve connectivity, with uncut and continuous-cut controls; not arbitrary free-form cutting.

Demo source and reports. The eight registered research experiences map into these same records. In particular, the 30 fixed-parameter mass/size-transfer conditions produce one engineering pass and no joint transient pass; that is a parameter-transfer result, not proof that the engine cannot perform the tasks.

A separate output-lifecycle defect is tracked by #243: reusing the apple default output directory can mix a new manifest with an old success summary. This is an evidence-integrity risk, not a solver failure. No contamination of the published records reviewed here has been established; the issue remains open.