RLWRLD · joocjun · Sep 16, 2026
diff --git a/docs/rlwrld/newton/deformable_apple_strip.png b/docs/rlwrld/newton/deformable_apple_strip.png
diff --git a/docs/rlwrld/newton/deformable_plum_strip.png b/docs/rlwrld/newton/deformable_plum_strip.png
diff --git a/docs/rlwrld/newton/deformable_polybag_cotton_strip.png b/docs/rlwrld/newton/deformable_polybag_cotton_strip.png
diff --git a/docs/rlwrld/newton/dragon_spoon_19cm_hull_fail.png b/docs/rlwrld/newton/dragon_spoon_19cm_hull_fail.png
diff --git a/docs/rlwrld/newton/dragon_spoon_19cm_variant_pass.png b/docs/rlwrld/newton/dragon_spoon_19cm_variant_pass.png
diff --git a/docs/rlwrld/newton/godmiddag_bowl_16_hull_fail.png b/docs/rlwrld/newton/godmiddag_bowl_16_hull_fail.png
diff --git a/docs/rlwrld/newton/godmiddag_bowl_16_variant_pass.png b/docs/rlwrld/newton/godmiddag_bowl_16_variant_pass.png
diff --git a/docs/rlwrld/newton/harmynta_deep_plate_22_hull_fail.png b/docs/rlwrld/newton/harmynta_deep_plate_22_hull_fail.png
diff --git a/docs/rlwrld/newton/harmynta_deep_plate_22_variant_pass.png b/docs/rlwrld/newton/harmynta_deep_plate_22_variant_pass.png
diff --git a/docs/rlwrld/newton/hex_nut_m20_variant_pass.png b/docs/rlwrld/newton/hex_nut_m20_variant_pass.png
diff --git a/docs/rlwrld/newton/ikea_365_bowl_rounded_16_hull_fail.png b/docs/rlwrld/newton/ikea_365_bowl_rounded_16_hull_fail.png
diff --git a/docs/rlwrld/newton/ikea_365_bowl_rounded_16_variant_pass.png b/docs/rlwrld/newton/ikea_365_bowl_rounded_16_variant_pass.png
diff --git a/docs/rlwrld/newton/sus304_flat_bar_fail.png b/docs/rlwrld/newton/sus304_flat_bar_fail.png
diff --git a/docs/rlwrld/newton_conformance.md b/docs/rlwrld/newton_conformance.md
@@ -0,0 +1,102 @@
+# SimReady props on Newton 1.5.2: conformance runner, findings and collision variants
+
+RLWRLD addition (2026-09-16), tools in `nv_core/testing_tools/rlwrld-newton-conformance/`.
+
+## Why
+
+The SimReady runtime tests only exist for Isaac Sim / PhysX. DexBench runs its scenes on Newton as well,
+and a package that passes `Prop-Robotics-Neutral 2.0.0` and the PhysX grasp-and-lift can still be
+unusable there: MuJoCo-Warp collides every mesh as one 64-vertex convex hull, so a bowl is a lid, a
+`convexDecomposition` depends on CoACD at every load, and a planar piece is refused. We needed the same
+three runtime tests on Newton, for all 171 handled props, to know which packages hold and what the
+engine needs that PhysX did not.
+
+| `godmiddag_bowl_16`, package colliders (one hull): the pads close inside the "lid" and tip the bowl | the same bowl with its Newton variant: pinched at the rim, lifted, shaken, set down |
+|---|---|
+| ![](newton/godmiddag_bowl_16_hull_fail.png) | ![](newton/godmiddag_bowl_16_variant_pass.png) |
+
+## What
+
+`newton15_cert.py` mirrors the three NVIDIA runtime tests on the standalone Newton 1.5.2 wheel: parse
+with Newton's USD importer, drop 1 cm onto a floor, drop onto a 15° walled slope, and grasp-and-lift on the
+authored `grasp_identifier_01` with a two-pad gantry (close, lift by max(0.3 m, 2 × longest edge), hold 1 s,
+shake 1 cm at 2 Hz for 1.5 s, hold, open). Verdict messages follow NVIDIA's wording so PhysX and Newton
+results can be compared line by line. Every asset runs in its own subprocess; a crash is recorded, not fatal.
+`newton15_cert_report.py` writes the comparison report. `make_newton_variant.py` authors the per-package
+Newton collision layer described below.
+
+## Results on the 171 handled props
+
+| test | pass | fail |
+|---|---|---|
+| parse (Newton importer) | 171 | 0 |
+| ground drop | 171 | 0 |
+| slope drop (15°, walled) | 170 | 1 |
+| grasp-and-lift, line on the body, tuned contact settings | 152 | 19 |
+| grasp-and-lift, line in the stage frame | 131 | 40 |
+| grasp-and-lift, stock Newton contact settings | 0 | 171 |
+
+Twenty-six packages disagree with PhysX. The ones PhysX passes and Newton fails are the `sdf` tableware
+(five bowls and plates), thin cutlery, a spark plug and a roundline cylinder that tip or roll off their
+hull, and a chips bag. The ones Newton passes and PhysX fails are props whose floor-standing PhysX verdict
+was a rig limit (the split shaft collar, several thin parts).
+
+| `dragon_spoon_19cm` on its hull: the pads close on nothing | with the variant: held through the shake |
+|---|---|
+| ![](newton/dragon_spoon_19cm_hull_fail.png) | ![](newton/dragon_spoon_19cm_variant_pass.png) |
+
+| `harmynta_deep_plate_22` on its hull: dropped at lift | with the variant: held |
+|---|---|
+| ![](newton/harmynta_deep_plate_22_hull_fail.png) | ![](newton/harmynta_deep_plate_22_variant_pass.png) |
+
+## What Newton needs from an asset that PhysX did not
+
+1. Every collider becomes convex. `sdf` and unauthored mesh colliders silently turn into one 64-vertex hull.
+2. No zero-thickness or sliver pieces: 1002 pieces in 16 packages had to be thickened to 1 mm by the runner.
+3. CoACD is a hard dependency for `convexDecomposition`, and it runs on every load.
+4. Contact stiffness is mass-scaled and the stock settings cannot hold a pinch grasp (0 of 171); the runner uses solref 4 ms, an elliptic cone, `impratio` 10, `condim 4` on the pads.
+5. Torsional friction is off by default; without `condim 4` on the gripper every off-centre pinch spins out.
+6. Soft contacts creep: props on the 15° slope slide about 1 cm/s and held objects creep millimetres in the jaws, so rest and hold checks need tolerances.
+7. Authored poses are not always rest poses (seven packages tip when dropped 1 cm), so the grasp line has to ride with the body.
+8. Mass, inertia and bound physics materials are honoured; multi-body vendor packages import as articulations.
+9. Hulls are capped at 64 vertices, coarser than PhysX: cylinders facet, rims flatten.
+10. `UsdPhysics.LoadUsdPhysicsFromRange` races on a body with many collision prims (heap corruption in about one load in four); `PXR_WORK_THREAD_LIMIT=1` before opening the stage avoids it.
+
+## Newton collision variants
+
+`make_newton_variant.py` writes `usd/variants/<name>_newton.usd` inside every package whose colliders the
+importer cannot honour: a flattened, self-contained copy in which each such collider is replaced by
+pre-decomposed convex pieces (CoACD with the importer's own light search, merge on, at most 32 hulls per
+collider and 64 vertices per hull, pieces thinner than 1 mm extruded to 1 mm), placed under the rigid body
+in the body's frame with the original physics material and a volume-split mass. The original mesh keeps
+rendering. Each variant is parsed back through Newton's importer in a fresh interpreter before it is
+accepted; the sidecar records the pieces, the CoACD settings and the certification verdict.
+
+| | package USD | Newton variant |
+|---|---:|---:|
+| colliders the importer replaced / thickened | 87 / 1002 | 0 / 0 |
+| parse time, all 85 | 95 s | 43 s |
+| ground drop pass | 85 | 85 |
+| slope drop pass | 85 | 83 |
+| grasp holds, line on the body | 74 | 79 |
+
+| `ikea_365_bowl_rounded_16` hull vs variant | `hex_nut_m20` variant (32 pieces) | `sus304_flat_bar`: a 3 mm sheet the pads cannot pinch on any engine |
+|---|---|---|
+| ![](newton/ikea_365_bowl_rounded_16_hull_fail.png) ![](newton/ikea_365_bowl_rounded_16_variant_pass.png) | ![](newton/hex_nut_m20_variant_pass.png) | ![](newton/sus304_flat_bar_fail.png) |
+
+## Deformables (examples)
+
+SimReady has no runtime test for deformable packages. `examples/newton15_soft_grasp.py` (tetrahedral fruit)
+and `examples/newton15_polybag_grasp.py` (cloth film with seal springs) run the same pad gantry on Newton's
+VBD soft-body solver: settle, close by a set squeeze, lift, hold, shake, open. The SpaceAI apple, plum and
+strawberry and the two polybags hold and deform plausibly.
+
+| apple | plum | polybag with cotton |
+|---|---|---|
+| ![](newton/deformable_apple_strip.png) | ![](newton/deformable_plum_strip.png) | ![](newton/deformable_polybag_cotton_strip.png) |
+
+## What it is not
+
+Not an engine plugin for `simready-benchmark`; a standalone runner that mirrors the tests' procedure so
+verdicts can be compared. Not a SimReady requirement: a package's profile pass stays a PhysX/Kit statement,
+and the Newton verdicts are recorded beside it.
diff --git a/nv_core/testing_tools/rlwrld-newton-conformance/README.md b/nv_core/testing_tools/rlwrld-newton-conformance/README.md
@@ -0,0 +1,46 @@
+# RLWRLD Newton conformance tools
+
+The SimReady runtime tests (FET003 drops, FET005 grasp-and-lift) run on Isaac Sim / PhysX. DexBench
+also runs its scenes on Newton, so we needed the same three tests on Newton to learn which
+`Prop-Robotics-Neutral` packages hold there and what the engine needs from an asset that PhysX did
+not. Everything here runs on the standalone Newton 1.5.2 wheel (MuJoCo-Warp rigid solver, warp 1.17),
+not inside Kit.
+
+| file | what |
+|---|---|
+| `newton15_cert.py` | The runner. Per package: parse with Newton's USD importer, ground drop (1 cm), slope drop (15°, walled), grasp-and-lift on the authored `grasp_identifier_01` with a two-pad gantry (close, lift, hold 1 s, shake 1 cm at 2 Hz, hold, open). Verdict wording follows NVIDIA's. Driver mode runs every asset in its own subprocess (a crash is a finding, not the end of the run). |
+| `newton15_cert_report.py` | Turns the merged `results.json` (plus the PhysX verdicts) into the report: totals, Newton-vs-PhysX disagreements, drop tests, collider approximations the importer could not honour. |
+| `make_newton_variant.py` | Authors `usd/variants/<name>_newton.usd` inside a package: a flattened copy whose `sdf`, `convexDecomposition` and unauthored mesh colliders become pre-decomposed convex pieces (CoACD, ≤ 32 hulls per collider, ≤ 64 vertices per hull, pieces thinner than 1 mm extruded to 1 mm), parsed back through Newton's importer before it is accepted. Records itself in the package sidecar and changelog (`rlwrld_sidecar.py`). |
+| `examples/newton15_soft_grasp.py`, `examples/newton15_polybag_grasp.py` | Grasp-and-deform demos for tetrahedral (fruit) and cloth (polybag) packages on Newton's VBD soft-body solver: the same pad gantry, on assets SimReady has no runtime test for. |
+
+## Running
+
+```bash
+python -m venv newton15 && newton15/bin/pip install "newton==1.5.2" "newton[importers]" usd-core imageio[ffmpeg] pyglet
+# the three tests over a manifest of package USDs (one path per line, relative to --assets-root)
+newton15/bin/python newton15_cert.py --assets-root <pool> --manifest packages.txt --out out/ --tests parse,ground_drop,slope_drop,grasp_and_lift
+# the report, against the PhysX verdicts of the same packages
+newton15/bin/python newton15_cert_report.py --main out/results.json --physx grasp_verdicts.json --fet003 fet003_results.json --out report.md
+# Newton collision variants for the packages the importer cannot honour
+newton15/bin/python make_newton_variant.py --assets-root <pool> --candidates candidates.json --write
+```
+
+`PXR_WORK_THREAD_LIMIT=1` is set by the runner before any stage opens: `UsdPhysics.LoadUsdPhysicsFromRange`
+races on a rigid body with many collision prims (heap corruption in about one load of four with usd-core
+26.3 on a 32-piece body; never with the work pool single-threaded).
+
+## What it found on 171 packages (2026-09-16)
+
+| test | pass / 171 |
+|---|---|
+| parse | 171 |
+| ground drop | 171 |
+| slope drop | 170 (a curved single-hull fork rocks and creeps) |
+| grasp-and-lift, line on the body | 152 |
+| grasp-and-lift, stock Newton contact settings | 0 |
+
+The 19 grasp failures were almost all colliders: MuJoCo-Warp reduces every mesh collider to one 64-vertex
+convex hull, so an `sdf` bowl becomes a lid and a `convexDecomposition` depends on CoACD running at load.
+With the Newton variants (85 packages, 2460 pieces) the bowls, plates and the dragon cutlery hold, grasp
+holds go from 74 to 79 of the 85, and the importer replaces or thickens nothing. The full write-up with
+frame strips is `docs/rlwrld/newton_conformance.md`.