newton/tests/test_multiworld_body_properties.py (1)

114-121: Remove unused variable bodies_per_world in the reference loop.

Line 117 computes bodies_per_world but it's never used within the loop. For single-world models, direct indexing with body_mass[0] is correct.

♻️ Remove unused variable

         for mass_val, label in [(mass_a, "A"), (mass_b, "B")]:
             model = self._build_articulated_model(num_worlds=1, device=device)
-            bodies_per_world = model.body_count // model.world_count
             body_mass = model.body_mass.numpy()
             body_mass[0] = mass_val  # base body
             model.body_mass.assign(body_mass)
             ref_q[label] = self._run_sim(model)

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@newton/tests/test_multiworld_body_properties.py` around lines 114 - 121,
Remove the unused local variable bodies_per_world from the reference loop in
test_multiworld_body_properties.py: inside the loop that builds the model (see
_build_articulated_model and use of model.body_mass), delete the computation
bodies_per_world = model.body_count // model.world_count since it is never used;
leave the rest of the loop (setting body_mass[0] and
model.body_mass.assign(...), and calling self._run_sim(model) to populate ref_q)
unchanged.

newton/_src/solvers/mujoco/kernels.py (1)

578-580: Fix is correct; consider extracting the global index for consistency with the sibling kernel.

The change correctly uses the global joint index for joint_type and joint_child lookups, aligning with how joint_q_start and joint_qd_start are already indexed.

For consistency with convert_mj_coords_to_warp_kernel (line 453), consider precomputing the global index once:

♻️ Optional: Extract global index to match sibling kernel pattern

     worldid, jntid = wp.tid()

+    joint_id = joints_per_world * worldid + jntid
+
     # Skip loop joints — they have no MuJoCo qpos/qvel entries
     q_i = mj_q_start[jntid]
     if q_i < 0:
         return

     qd_i = mj_qd_start[jntid]
-    type = joint_type[joints_per_world * worldid + jntid]
-    wq_i = joint_q_start[joints_per_world * worldid + jntid]
-    wqd_i = joint_qd_start[joints_per_world * worldid + jntid]
+    type = joint_type[joint_id]
+    wq_i = joint_q_start[joint_id]
+    wqd_i = joint_qd_start[joint_id]

And at line 610:

         # Get CoM offset in world frame
-        child = joint_child[joints_per_world * worldid + jntid]
+        child = joint_child[joint_id]

Also applies to: 610-611

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@newton/_src/solvers/mujoco/kernels.py` around lines 578 - 580, Extract the
global joint index once (e.g., compute a local variable like global_jnt =
joints_per_world * worldid + jntid) and use that variable for all array lookups
instead of repeating the expression; update usages in the kernel where
joint_type, joint_child, joint_q_start, and joint_qd_start are accessed (the
occurrences around the current block and the similar lookups at the later spots
around the 610-611 area) so indexing is consistent with
convert_mj_coords_to_warp_kernel and avoids duplicated arithmetic.

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

newton/tests/test_multiworld_body_properties.py (2)

1-1: ⚠️ Potential issue | 🟡 Minor

Update the SPDX header to the 2026 creation year.

This file is new in the current PR, so the header should use 2026, not 2025.

As per coding guidelines, “In SPDX copyright lines, use the year the file was first created. Do not create date ranges or update the year when modifying a file.”

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@newton/tests/test_multiworld_body_properties.py` at line 1, Update the SPDX
copyright header line in this new file by changing the creation year from 2025
to 2026: modify the SPDX line (the top-of-file comment starting with
"SPDX-FileCopyrightText") to use "2026" and do not create a date range; ensure
no other SPDX lines or ranges are added.

---

134-145: ⚠️ Potential issue | 🟡 Minor

Make the joint_q comparisons quaternion-sign invariant.

These assertions still compare the free-joint quaternion verbatim. q and -q encode the same pose, so this regression can fail even when the multi-world and single-world trajectories match. Please route them through a helper that flips one side when the quaternion dot product is negative.

Based on learnings, “Quaternions in builder/test assertions use Warp’s XYZW convention (identity [0, 0, 0, 1]). In newton/tests/test_import_mjcf.py (e.g., test_fit_box_to_mesh_inertia_rotated), prefer identity/non-identity checks that account for q and -q equivalence (e.g., via np.abs on components).”

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@newton/tests/test_multiworld_body_properties.py` around lines 134 - 145, The
quaternion comparisons between multi_q[...] and ref_q[...] need to be made
sign-invariant: implement or call a helper (e.g.,
align_quaternion_sign(quat_ref, quat_test) or flip_quat_if_needed) that computes
the dot product between the two 4D quaternions and multiplies one by -1 when the
dot is negative, then use the aligned quaternions in the
np.testing.assert_allclose calls for multi_q[0] vs ref_q["A"] and multi_q[1] vs
ref_q["B"] so q and -q are treated as equivalent.