deps: unisim-core 1.4.0 and mjbatch-uni 0.2.1

docs/sphinx/source/changelog.md

@@ -67,6 +67,29 @@ PyPI 版本变更与未发布变更；发布日期采用 PyPI 上传日期。完
   Reward manager 不再在 reset extras 中重复输出 `Episode_Reward/*`；既有
   episode log 仍是这些指标的数据源（#1570）。
 
+### Changed / 变更
+
+- The MuJoCo executor dependency is pinned to `mjbatch-uni~=0.2.1` and
+  `unisim-core` moves to `>=1.4.0`
+  ([unisim#71](https://github.com/unilabsim/unisim/issues/71),
+  [mjbatch_uni#28](https://github.com/unilabsim/mjbatch_uni/issues/28)).
+  unisim-core 1.4.0 adds per-environment gravity reset, the cross-backend
+  body wrench/torque contract, and per-substep callback wrenches through
+  `PreStepControlOutput`; on the MuJoCo pre-step control path the executor's
+  split-substep sensor copyout now refreshes tracked body state at every
+  substep boundary (about 14x faster than the previous host-side recompute,
+  with bit-identical trajectories). Tests that pinned the previous executor
+  callback protocol and Isaac worker diagnostic text were updated to the
+  published contracts.
+  MuJoCo 执行器依赖固定为 `mjbatch-uni~=0.2.1`，`unisim-core` 升级到
+  `>=1.4.0`（unisim#71、mjbatch_uni#28）。unisim-core 1.4.0 新增逐环境
+  重力 reset、跨后端 body wrench/torque 契约，以及通过
+  `PreStepControlOutput` 的逐子步 callback wrench；MuJoCo pre-step control
+  路径改用执行器的 split-substep 传感器增量拷出，在每个子步边界刷新
+  tracked body state（相比此前主机端重算约 14 倍加速，轨迹逐位一致）。
+  原先固定旧执行器 callback 协议与 Isaac worker 诊断文案的测试已更新到
+  已发布契约。
+
 ## 1.2.0 (2026-09-10)
 
 ### Breaking changes / 破坏性变更

pyproject.rocm.toml

@@ -24,9 +24,10 @@ requires-python = ">=3.10,<3.14"
 dependencies = [
     "numpy",
     # Physics implementations are provided by the independently released
-    # unisim-core package. The 1.3.0 release carries the fixed model variant
-    # and per-world reset-default contracts.
-    "unisim-core>=1.3.0",
+    # unisim-core package. The 1.4.0 release carries the fixed model variant,
+    # per-world reset-default, per-env gravity, and substep body-wrench
+    # contracts.
+    "unisim-core>=1.4.0",
     # RL algorithms and async runtimes live in the independently released
     # uni-rl package (distribution name ``unilab-rl``); see pyproject.toml.
     "unilab-rl==1.2.0",
@@ -80,7 +81,7 @@ mujoco = [
     # time, so isolated builds are correct and no compiler preflight is
     # needed.
     "mujoco~=3.11.0",
-    "mjbatch-uni~=0.2.0",
+    "mjbatch-uni~=0.2.1",
 ]
 motrix = ["motrixsim-core==0.8.2"]
 viser = ["viser>=1.0.26", "trimesh>=3.21.7"]

pyproject.toml

@@ -41,9 +41,10 @@ requires-python = ">=3.10,<3.14"
 dependencies = [
     "numpy",
     # Physics implementations are provided by the independently released
-    # unisim-core package. The 1.3.0 release carries the fixed model variant
-    # and per-world reset-default contracts.
-    "unisim-core>=1.3.0",
+    # unisim-core package. The 1.4.0 release carries the fixed model variant,
+    # per-world reset-default, per-env gravity, and substep body-wrench
+    # contracts.
+    "unisim-core>=1.4.0",
     # RL algorithms and async runtimes (PPO/APPO/SAC/TD3 runners,
     # collectors, IPC, logging) live in the independently released uni-rl
     # package (distribution name ``unilab-rl``), consumed via the injected
@@ -122,7 +123,7 @@ mujoco = [
     # not a UniLab config change.
     "mujoco~=3.11.0",
     # The batch engine is the published unilabsim mjbatch fork.
-    "mjbatch-uni~=0.2.0",
+    "mjbatch-uni~=0.2.1",
 ]
 mjwarp = [
     # Keep the Warp backend on the same MuJoCo minor line as the host backend.
@@ -172,7 +173,7 @@ viser = ["viser>=1.0.26", "trimesh>=3.21.7"]
 # required-environments; elsewhere the extra is empty and the CLI reports a
 # targeted runtime diagnostic.
 superdex = [
-    "unisim-core[superdex]>=1.3.0 ; python_version >= '3.12' and sys_platform == 'linux' and platform_machine == 'x86_64'",
+    "unisim-core[superdex]>=1.4.0 ; python_version >= '3.12' and sys_platform == 'linux' and platform_machine == 'x86_64'",
 ]
 
 [dependency-groups]

tests/base/test_backend_pre_step_control.py

@@ -7,8 +7,17 @@
 from unisim.backend.base import SimBackend
 
 
+def _contract_backend(fn=None):
+    """A minimal owner object exposing the pre-step-control contract methods."""
+    backend = SimpleNamespace(_pre_step_control_fn=fn)
+    backend._convert_pre_step_control = lambda ctrl: SimBackend._convert_pre_step_control(
+        backend, ctrl
+    )
+    return backend
+
+
 def test_pre_step_control_default_noop() -> None:
-    backend = SimpleNamespace(_pre_step_control_fn=None)
+    backend = _contract_backend()
     ctrl = np.zeros((2, 3), dtype=np.float32)
 
     out = SimBackend._apply_pre_step_control(backend, ctrl)  # type: ignore[arg-type]
@@ -17,7 +26,7 @@ def test_pre_step_control_default_noop() -> None:
 
 
 def test_pre_step_control_applies_registered_converter() -> None:
-    backend = SimpleNamespace(_pre_step_control_fn=None)
+    backend = _contract_backend()
     ctrl = np.array([[1.0, 2.0, 3.0]], dtype=np.float32)
 
     SimBackend.set_pre_step_control(  # type: ignore[arg-type]
@@ -34,7 +43,7 @@ def test_pre_step_control_applies_registered_converter() -> None:
 
 
 def test_pre_step_control_rejects_shape_mismatch() -> None:
-    backend = SimpleNamespace(_pre_step_control_fn=lambda current_backend, ctrl: ctrl[:, :1])
+    backend = _contract_backend(lambda current_backend, ctrl: ctrl[:, :1])
     ctrl = np.zeros((2, 3), dtype=np.float32)
 
     with pytest.raises(ValueError, match="pre-step control must return shape"):
@@ -70,6 +79,7 @@ def __init__(self, nq: int = 1, nv: int = 1, nu: int = 2, nbody: int = 1) -> Non
         }
         self.step_calls: list[dict] = []
         self.callback_controls: list[np.ndarray] = []
+        self.callback_channels: list[np.ndarray] = []
 
     def bind(self, name: str, dtype=None) -> np.ndarray:
         return self._bufs[name]
@@ -81,19 +91,28 @@ def step(
         history=None,
         *,
         callback=None,
+        substep_sensor_copyout=None,
     ) -> None:
         self.step_calls.append(
             {
                 "nstep": nstep,
                 "callback": callback,
+                "substep_sensor_copyout": substep_sensor_copyout,
             }
         )
         state = self._state
         ctrl_buf = self._bufs["ctrl"]
         for k in range(nstep):
             if callback is not None:
-                callback(k, state, ctrl_buf)
+                if substep_sensor_copyout is None:
+                    callback(k, state, ctrl_buf)
+                else:
+                    start, stop = substep_sensor_copyout
+                    callback(k, state, ctrl_buf, self._bufs["sensordata"][:, start:stop])
                 self.callback_controls.append(ctrl_buf.copy())
+                # The persistent wrench channel as the substep's physics
+                # sees it: composed after the callback, before integration.
+                self.callback_channels.append(self._bufs["xfrc_applied"].copy())
             state += 1.0
         # End-of-call copy-out of the bound input/derived fields.
         self._bufs["qpos"][:] = state[:, self.qpos_slice]
@@ -118,6 +137,7 @@ def _fake_mujoco_backend(pre_step_control_fn=None):
     pool = _FakeMjBatch()
     backend = object.__new__(MuJoCoBackend)
     backend._pre_step_control_fn = pre_step_control_fn
+    backend._tracked_sensor_copyout_range = None
     backend._num_envs = 1
     backend._np_dtype = np.float32
     backend.nq = pool._nq
@@ -219,7 +239,7 @@ def test_mujoco_step_writes_xfrc_absolutely_and_clears_pending() -> None:
     np.testing.assert_allclose(backend._pending_xfrc_applied, [[0.0] * 6])
 
 
-def test_mujoco_step_with_pre_step_control_stages_xfrc_before_dispatch() -> None:
+def test_mujoco_step_with_pre_step_control_applies_staged_xfrc_each_substep() -> None:
     backend = _fake_mujoco_backend()
     backend._pending_xfrc_applied = np.full((1, 6), 7.0, dtype=np.float64)
 
@@ -230,11 +250,18 @@ def test_mujoco_step_with_pre_step_control_stages_xfrc_before_dispatch() -> None
 
     pool = backend._pool
     assert len(pool.step_calls) == 1
-    # The callback protocol writes ctrl only; the wrench rides its own channel.
+    # The staged wrench is recomposed onto the persistent channel before every
+    # substep; each recorded channel is the wrench that substep's physics
+    # actually consumed.
     assert len(pool.callback_controls) == 2
     for cb_control in pool.callback_controls:
         np.testing.assert_allclose(cb_control, [[1.5, 0.5]])
-    np.testing.assert_allclose(backend._xfrc_view.reshape(1, -1), [[7.0] * 6])
+    assert len(pool.callback_channels) == 2
+    for channel in pool.callback_channels:
+        np.testing.assert_allclose(channel.reshape(1, -1), [[7.0] * 6])
+    # The call finishes by clearing the persistent channel and the staging
+    # buffer so nothing leaks into the next control step.
+    np.testing.assert_allclose(backend._xfrc_view.reshape(1, -1), [[0.0] * 6])
     np.testing.assert_allclose(backend._pending_xfrc_applied, [[0.0] * 6])
 
 

tests/base/test_isaacgym_backend.py

@@ -252,7 +252,9 @@ def test_dependencies_missing_runtime_is_actionable(
 ) -> None:
     monkeypatch.setenv(ENV_HOME, str(tmp_path / "empty"))
     monkeypatch.delenv(ENV_PYTHON, raising=False)
-    with pytest.raises(IsaacGymDependencyError, match="setup_isaacgym_env.sh"):
+    with pytest.raises(
+        IsaacGymDependencyError, match="could not find the Python 3.8 worker interpreter"
+    ):
         resolve_isaacgym_runtime()
 
 

tests/base/test_isaacsim_backend.py

@@ -184,7 +184,9 @@ def test_dependencies_override_and_missing_layout(
     empty = tmp_path / "empty"
     monkeypatch.setenv(ENV_HOME, str(empty))
     monkeypatch.delenv(ENV_PYTHON, raising=False)
-    with pytest.raises(IsaacSimDependencyError, match="setup_isaacsim_env.sh"):
+    with pytest.raises(
+        IsaacSimDependencyError, match="could not find the Python 3.11 worker interpreter"
+    ):
         resolve_isaacsim_runtime()