[Bump] Bump Newton to 1.6.0rc1 for newton#4017 (#7445)

## Summary

- Bumps Newton from `1.5.1` to `1.6.0rc1`, the first release carrying
[newton#4017](newton-physics/newton#4017),
unblocking the MuJoCo tendon adapter that PR #7161 needs.
- `warp-lang` and `mujoco`/`mujoco-warp` move with it, because
`1.6.0rc1` requires them.
- Pins `newton[sim]` in `[project].dependencies` as well, so a wheel
built from this metadata resolves the fix rather than falling back to
the newest stable.
- Updates test literals in `isaaclab` and `isaaclab_newton` for Newton's
new actuator class names, and asks for a determinism guarantee in the
manipulation determinism test.

# Description

`resolve_fixed_tendon_actuator_columns` reads `mujoco.actuator_trntype`
through an
`ArticulationView`. Every Newton 1.5.x release rejects that with *"has
custom frequency
'mujoco:actuator' which is not supported by ArticulationView"*;
newton#4017 makes
`ArticulationView` generic over custom frequencies.

`1.6.0rc1` (published 2026-09-03) is the first release containing that
fix — verified two ways:
the merge commit `4055a159` is an ancestor of `v1.6.0rc1`, and
`custom_frequency_label_attributes`,
the symbol the fix introduces in `newton/_src/sim/model.py`, is present
at `v1.6.0rc1` and absent
at `v1.5.1`. No `1.6.0` final exists yet, and no 1.5.x backport carries
the fix, so `1.6.0rc1` is
the only released option.

The release requires `warp-lang>=1.17.0` and `mujoco-warp~=3.12.0`, so
those pins move too:

| Package | Before | After |
| --- | --- | --- |
| `newton[sim]` | `1.5.1` | `1.6.0rc1` |
| `warp-lang` | `1.16.0` | `1.17.0` |
| `mujoco` | `3.11.0` | `3.12.0` |
| `mujoco-warp` | `3.11.0` | `3.12.0` |

`warp-lang` is pinned to the `1.17.0` release rather than left as
`>=1.17.0`: Newton's prerelease
enables prerelease resolution, and an unbounded bound then resolves to a
nightly
(`1.18.0.dev20260903`).

## Wheel installs get the pin too

`[tool.uv].override-dependencies` only reaches `uv sync`, so it does not
constrain a wheel built
from this metadata. `1.6.0rc1` is a prerelease, which a default resolve
skips, so a loose bound
resolved `newton 1.5.1` — without newton#4017, the fix this bump exists
to deliver. Since
`wheel.yml` builds on every push to `develop`, that would have shipped.

`[project].dependencies` now pins `newton[sim]==1.6.0rc1`, matching the
override. A default
resolve of the new metadata returns `1.6.0rc1` under both `uv pip
compile` and `pip download`,
with no `--pre` flag. `isaacsim-core==6.0.1.0` requires
`newton[sim]==1.2.1`, so installing the
`isaacsim` extra still needs an override — exactly as on `develop`,
which carries an equally
exact pin in the same place.

## Newton actuator class rename

`1.6.0rc1` introduces `DrivePD` / `DriveNeuralMLP` / `DriveNeuralLSTM`
alongside the existing
`ControllerPD` / `ControllerNeural*`. Both names remain importable, so
this is **not** a breaking
API change and no production code changes. Tests in `isaaclab` and
`isaaclab_newton` assert on the
class name Newton *constructs with*, which switched to the `Drive*`
form, so those literals are
updated.

## Determinism

`test_manipulation_env_determinism[cuda-Isaac-Open-Drawer-Franka]`
failed on this branch in CI.
Two separate questions: whether the bump caused it, and why it failed at
all.

**Not caused by the bump.** Each arm was run 5x on the same machine,
because a single run of an
intermittent test samples a coin rather than measuring it:

| GPU | develop pins (`newton 1.5.1` / `warp 1.16.0` / `mujoco 3.11.0`)
| this PR (`newton 1.6.0rc1` / `warp 1.17.0` / `mujoco 3.12.0`) |
| --- | --- | --- |
| RTX PRO 6000 Blackwell | 3 / 5 failed | 3 / 5 failed |
| L40 | 3 / 3 passed | 4 / 4 passed |

Both pin sets fail at the same rate on Blackwell and both are clean on
L40, so the failure is
hardware-dependent, not version-dependent. Intermediate Newton revisions
(the commits before and
at [newton#4119](newton-physics/newton#4119),
the MuJoCo 3.12 bump) show
the same behaviour, so neither Warp 1.17, MuJoCo 3.12, nor #4119 is
implicated.

**Why it failed.** The test asserted bit-reproducible rewards without
ever asking for a
determinism guarantee. Newton defaults to
`wp.DeterministicMode.NOT_GUARANTEED`, under which
Warp's atomics may accumulate in any order, so two runs of the same
environment are not required
to agree — the assertion was stricter than the configuration it ran
under, and whether it held
came down to GPU scheduling.

The Newton-backed case now passes `deterministic_mode="run_to_run"`,
which is exactly the
guarantee this test asserts: one device, one process, run twice.
`deterministic_mode` is a Newton
setting, so the PhysX-backed `IsaacContrib-Lift-Cube-Franka` case leaves
it unset and the helper
keeps asserting that the mode is only applied to a `NewtonCfg` backend.
This is a
test-configuration fix, not a masking of the intermittency: the
guarantee the assertion depends
on is now actually requested.

Related: [#7561](#7561)
addresses the adjacent layer and
does not overlap this PR. Newton's solvers take a per-module
`deterministic` argument, which is
what `deterministic_mode` reaches and what this PR sets; Newton's sensor
and geometry kernels take
no such argument and fall back to the `warp.config.deterministic`
global, which #7561 raises from
the same config field in the RL entrypoint.

## Type of change

- Bug fix (non-breaking change which fixes an issue)

## Release backport

- [ ] <!-- backport-active-release --> Backport this pull request to the
active release branch after it merges into `develop`

## Validation

- `uv lock --check` — clean; the relock moves exactly the four packages
above, and the
  `[project].dependencies` pin moves exactly one further lockfile line
- Default resolve of the published metadata returns `newton==1.6.0rc1`
under both
  `uv pip compile` and `pip download`, with no `--pre` flag
- `uv run --frozen python -m pytest
source/isaaclab/test/cli/test_wheel_builder_metadata.py
source/isaaclab/test/cli/test_uv_run_pyproject.py` — 21 passed, which
enforces that the two generated `uv-overrides.txt` files match the
override table and that `[tool.isaaclab.versions]` mirrors the
dependency pins
- `uv run isaaclab -f` — passed
- Full Docker + Tests suite — **48 passed / 0 failed** on `b184335e80b`

Author: hujc7

pyproject.toml

@@ -51,7 +51,7 @@ dependencies = [
     "pyglet>=2.1.6,<3",
     "transformers==4.57.6",
     "einops",
-    "warp-lang==1.16.0",
+    "warp-lang==1.17.0",
     "matplotlib>=3.10.3",
     # pillow: floor, not exact — an exact pin below Isaac Sim's prebundled version forces a
     # downgrade that deletes the prebundled copy other extensions symlink into (nvbugs 6410989).
@@ -87,8 +87,11 @@ dependencies = [
     "rsl-rl-lib==5.4.1",  # default RL framework
     "onnxscript>=0.5",
     # ----- newton (default physics engine) -----
-    # Pin the Newton build used by both workspace and wheel installs until the 1.6 release.
-    "newton[sim] @ git+https://github.com/newton-physics/newton.git@24bd863528d6b91137408930d0fbe8fa216ad962",
+    # Pin the Newton release used by both workspace and wheel installs until 1.6.0 final.
+    # [tool.uv].override-dependencies only reaches uv sync, so a wheel built from this
+    # metadata would otherwise resolve the newest stable (1.5.1) and miss newton#4017:
+    # 1.6.0rc1 is a prerelease, which a default resolve skips.
+    "newton[sim]==1.6.0rc1",
     # Import and mesh-processing packages used by Newton, including the ones that honoring
     # USD-authored ``physics:approximation`` requires. Keep these explicit instead of selecting
     # newton[importers], whose standalone USD dependency would overlap with usd-exchange.
@@ -222,7 +225,7 @@ torchaudio = "2.11.0"
 ovphysx = "0.5.11"
 ovrtx = "0.4.1.364340"
 ovstage = "0.1.1.355824"
-warp = "1.16.0"
+warp = "1.17.0"
 
 [tool.ruff]
 line-length = 120
@@ -402,9 +405,9 @@ environments = [
 # Torch routes through [tool.uv.sources]. Values mirror [tool.isaaclab.versions] where applicable.
 override-dependencies = [
     "numpy>=2",
-    "mujoco~=3.11.0",
-    "mujoco-warp~=3.11.0",
-    "newton[sim] @ git+https://github.com/newton-physics/newton.git@24bd863528d6b91137408930d0fbe8fa216ad962",
+    "mujoco~=3.12.0",
+    "mujoco-warp~=3.12.0",
+    "newton[sim]==1.6.0rc1",
     # Force the Newton-matched schemas over isaacsim's ==0.2.0 pin.
     "newton-usd-schemas>=0.4.1",
     "torch==2.11.0",

source/isaaclab/changelog.d/jichuanh-newton-pin-4017.minor.rst

@@ -0,0 +1,10 @@
+Changed
+^^^^^^^
+
+* Changed the pinned Newton version from ``1.5.1`` to ``1.6.0rc1``, the first release carrying
+  `newton#4017 <https://github.com/newton-physics/newton/pull/4017>`_, which makes
+  ``ArticulationView`` generic over custom frequencies. Reading a MuJoCo actuator attribute such
+  as ``mujoco.actuator_trntype`` through an articulation view previously raised *"has custom
+  frequency 'mujoco:actuator' which is not supported by ArticulationView"*.
+* Changed ``warp-lang`` from ``1.16.0`` to ``1.17.0`` and ``mujoco``/``mujoco-warp`` from
+  ``3.11.0`` to ``3.12.0``, which Newton ``1.6.0rc1`` requires.

source/isaaclab/test/actuators/test_newton_adapter.py

@@ -7,7 +7,7 @@
 
 import numpy as np
 import pytest
-from newton.actuators import ClampingDCMotor, ClampingMaxEffort, ClampingPositionBased, ControllerPD
+from newton.actuators import ClampingDCMotor, ClampingMaxEffort, ClampingPositionBased, DrivePD
 
 from pxr import Usd, UsdGeom, UsdPhysics
 
@@ -110,7 +110,7 @@ def test_from_usd_groups_by_structure_and_preserves_per_dof_values():
     assert len(actuators) == 4
 
     pd = next(actuator for actuator in actuators if [type(c) for c in actuator.clamping] == [ClampingMaxEffort])
-    assert type(pd.controller) is ControllerPD
+    assert type(pd.controller) is DrivePD
     np.testing.assert_array_equal(pd.indices.numpy(), [0, 1, 6, 7])
     np.testing.assert_allclose(pd.controller.kp.numpy(), [11.0, 22.0, 11.0, 22.0])
     np.testing.assert_allclose(pd.controller.kd.numpy(), [1.5, 2.5, 1.5, 2.5])
@@ -119,7 +119,7 @@ def test_from_usd_groups_by_structure_and_preserves_per_dof_values():
     assert pd.delay.buf_depth == 4
 
     dc = next(actuator for actuator in actuators if [type(c) for c in actuator.clamping] == [ClampingDCMotor])
-    assert type(dc.controller) is ControllerPD
+    assert type(dc.controller) is DrivePD
     assert dc.delay is None
     np.testing.assert_array_equal(dc.indices.numpy(), [2, 3, 8, 9])
     np.testing.assert_allclose(dc.controller.kp.numpy(), [33.0, 44.0, 33.0, 44.0])

source/isaaclab/test/install_ci/uv_pip/test_uv_pip_install_isaaclab_all_trains_cartpole.py

@@ -15,8 +15,8 @@
          Reinstall AFTER the wheel install: unsafe-best-match re-resolves torch from PyPI to CPU.)
     - (aarch64 only) export LD_PRELOAD=/lib/aarch64-linux-gnu/libgomp.so.1
 Tests:
-    - python -c "import importlib.metadata as m; assert m.version('newton') == '1.6.0.dev0'"
-        -> verify the wheel resolves the pinned Newton 1.6 development build
+    - python -c "import importlib.metadata as m; assert m.version('newton') == '1.6.0rc1'"
+        -> verify the wheel resolves the pinned Newton release
     - uv run isaaclab train --rl_library rsl_rl --task Isaac-Cartpole-Direct --num_envs 16
         presets=newton_mjwarp --max_iterations 5; uv run isaaclab train --rl_library rsl_rl
         --task Isaac-Cartpole-Camera-Direct --num_envs 16 presets=newton_mjwarp,newton_renderer --max_iterations 2
@@ -57,7 +57,7 @@ def test_uv_pip_install_isaaclab_all_trains_cartpole(self, isaaclab_root, wheel,
             assert result.returncode == 0, f"uv pip install {wheel}[all] failed:\n{result.stdout}\n{result.stderr}"
 
             result = self.run_in_uv_env(
-                ["python", "-c", "import importlib.metadata as m; assert m.version('newton') == '1.6.0.dev0'"],
+                ["python", "-c", "import importlib.metadata as m; assert m.version('newton') == '1.6.0rc1'"],
                 cwd=isaaclab_root,
             )
             assert result.returncode == 0, (

source/isaaclab/test/install_ci/uv_pip/uv-overrides.txt

@@ -1,7 +1,7 @@
 numpy>=2
-mujoco~=3.11.0
-mujoco-warp~=3.11.0
-newton[sim] @ git+https://github.com/newton-physics/newton.git@24bd863528d6b91137408930d0fbe8fa216ad962
+mujoco~=3.12.0
+mujoco-warp~=3.12.0
+newton[sim]==1.6.0rc1
 newton-usd-schemas>=0.4.1
 torch==2.11.0
 torchvision==0.26.0

source/isaaclab_newton/changelog.d/jichuanh-newton-actuator-rename-test.skip

@@ -0,0 +1,3 @@
+Test-only change: the actuator tests now assert Newton's new drive class names
+(``DrivePD``, ``DriveNeuralMLP``, ``DriveNeuralLSTM``). Newton 1.6.0rc1 still exports the
+old ``Controller*`` aliases, so there is no user-visible behavior change.

source/isaaclab_newton/test/assets/test_newton_actuators_newton.py

@@ -598,7 +598,7 @@ def test_has_delay(self):
 
     def test_controller_is_pd(self):
         for a in self.result["actuator_info"]:
-            self.assertEqual(a["controller_type"], "ControllerPD")
+            self.assertEqual(a["controller_type"], "DrivePD")
 
 
 # ---------------------------------------------------------------------------
@@ -772,7 +772,7 @@ def test_kfe_controller_is_pd(self):
         kfe_acts = [a for a in self.result["actuator_info"] if "ClampingPositionBased" in a["clamping_types"]]
         self.assertTrue(len(kfe_acts) > 0, "No actuator with position-based clamping found")
         for a in kfe_acts:
-            self.assertEqual(a["controller_type"], "ControllerPD")
+            self.assertEqual(a["controller_type"], "DrivePD")
 
     def test_kfe_has_position_based_clamping(self):
         kfe_acts = [a for a in self.result["actuator_info"] if "ClampingPositionBased" in a["clamping_types"]]
@@ -845,11 +845,11 @@ def test_num_actuators(self):
         self.assertGreaterEqual(self.result["num_actuators"], 2)
 
     def test_has_neural_mlp_controller(self):
-        mlp_acts = [a for a in self.result["actuator_info"] if a["controller_type"] == "ControllerNeuralMLP"]
+        mlp_acts = [a for a in self.result["actuator_info"] if a["controller_type"] == "DriveNeuralMLP"]
         self.assertTrue(len(mlp_acts) > 0, "No NeuralMLP controller found")
 
     def test_mlp_has_dc_motor_clamping(self):
-        mlp_acts = [a for a in self.result["actuator_info"] if a["controller_type"] == "ControllerNeuralMLP"]
+        mlp_acts = [a for a in self.result["actuator_info"] if a["controller_type"] == "DriveNeuralMLP"]
         for a in mlp_acts:
             self.assertIn("ClampingDCMotor", a["clamping_types"])
 
@@ -890,11 +890,11 @@ def test_num_actuators(self):
         self.assertGreaterEqual(self.result["num_actuators"], 2)
 
     def test_has_neural_lstm_controller(self):
-        lstm_acts = [a for a in self.result["actuator_info"] if a["controller_type"] == "ControllerNeuralLSTM"]
+        lstm_acts = [a for a in self.result["actuator_info"] if a["controller_type"] == "DriveNeuralLSTM"]
         self.assertTrue(len(lstm_acts) > 0, "No NeuralLSTM controller found")
 
     def test_lstm_has_dc_motor_clamping(self):
-        lstm_acts = [a for a in self.result["actuator_info"] if a["controller_type"] == "ControllerNeuralLSTM"]
+        lstm_acts = [a for a in self.result["actuator_info"] if a["controller_type"] == "DriveNeuralLSTM"]
         for a in lstm_acts:
             self.assertIn("ClampingDCMotor", a["clamping_types"])
 

source/isaaclab_tasks/changelog.d/jichuanh-manipulation-determinism-mode.rst

@@ -0,0 +1,8 @@
+Fixed
+^^^^^
+
+* Fixed ``test_manipulation_env_determinism`` asserting bit-reproducible rewards without requesting
+  a determinism guarantee. Newton defaults to ``wp.DeterministicMode.NOT_GUARANTEED``, under which
+  Warp's atomics may accumulate in any order, so the test failed intermittently depending on GPU
+  scheduling. It now passes ``deterministic_mode="run_to_run"``, as the Newton cartpole cases
+  already did.

source/isaaclab_tasks/test/core/test_environment_determinism.py

@@ -36,16 +36,20 @@ def setup_environment():
 
 
 @pytest.mark.parametrize(
-    "task_name",
+    ("task_name", "deterministic_mode"),
     [
-        "Isaac-Open-Drawer-Franka",
-        "IsaacContrib-Lift-Cube-Franka",
+        # Newton defaults to ``wp.DeterministicMode.NOT_GUARANTEED``, under which Warp's atomics may
+        # accumulate in any order, so two runs are not bit-reproducible. ``run_to_run`` is exactly
+        # the guarantee this test asserts: one device, one process, run twice. ``deterministic_mode``
+        # is a Newton setting, so the PhysX-backed task leaves it unset.
+        ("Isaac-Open-Drawer-Franka", "run_to_run"),
+        ("IsaacContrib-Lift-Cube-Franka", None),
     ],
 )
 @pytest.mark.parametrize("device", ["cuda", "cpu"])
-def test_manipulation_env_determinism(task_name, device):
+def test_manipulation_env_determinism(task_name, deterministic_mode, device):
     """Check deterministic environment creation for manipulation."""
-    _test_environment_determinism(task_name, device)
+    _test_environment_determinism(task_name, device, deterministic_mode=deterministic_mode)
 
 
 @pytest.mark.parametrize(

tools/wheel_builder/uv-overrides.txt

@@ -1,7 +1,7 @@
 numpy>=2
-mujoco~=3.11.0
-mujoco-warp~=3.11.0
-newton[sim] @ git+https://github.com/newton-physics/newton.git@24bd863528d6b91137408930d0fbe8fa216ad962
+mujoco~=3.12.0
+mujoco-warp~=3.12.0
+newton[sim]==1.6.0rc1
 newton-usd-schemas>=0.4.1
 torch==2.11.0
 torchvision==0.26.0