Search — dx

wp.func_grad raises KeyError 'output_arch' when registering gradients for functions using tile_matmul. The dummy ModuleBuilder used for return-type probing lacks an option that tile builtins expect.

tile_matmul silently computes wrong results on CUDA when given a strided tile_view operand. The cuBLASDx path derives leading dimensions from shape and ignores strides, unlike the CPU fallback.

Request to remove a SolverMuJoCo temporary COM-offset workaround once MuJoCo supports disabling simple-body optimization via a supported setting. The workaround currently adjusts body_ipos to force general mass-matrix storage.

Request to remove SolverMuJoCo line-search block-dimension override once upstream caps the block dimension. The workaround exists due to prior CUDA register limit failures.

Newton has inconsistent optional-output semantics across dynamics evaluation APIs where None can mean 'allocate and return' in one function and 'skip computing' in another. Request is to unify behavior before stabilizing the API family.

Nsight Systems on Orin Nano shows GA10B GPU DRAM bandwidth metrics as all zeros. User is unsure if this is a limitation or misconfiguration, blocking performance analysis.

In IsaacLab 3.0, camera data position remains static even when the robot arm moves, despite update_latest_camera_pose=True and attachment to the prim path. Users can see the camera move in the GUI but the queried tensors do not update, breaking programmatic perception.

Newton plans to deprecate `MeshAdjacency.add_edge` and `MeshAdjacency.edges`, moving away from dict-based representations toward a more efficient vectorized form. The issue tracks the deprecation timeline (warning in 1.4; deprecate in 1.6).

Useful RL script utilities live in `scripts/reinforcement_learning/common.py` but are not easily importable for downstream use. The request is to move it into a module or add initialization so it can be imported directly.

Newton has viewers and a `ViewerRerun`, but tests/apps must manually wire together model/state/contacts/control/solver data for coherent debugging. The request is for an opt-in standard Rerun-based debug session helper to simplify reproduction and inspection.

Newton reviews increasingly involve AI agents, and the project wants clearer, practical design/style guidance for both humans and agents. The issue proposes turning existing draft specs/audits into actionable review checklists to catch problems earlier.

MuJoCo’s MjSpec.to_xml appears to change joint/state-vector ordering after an attach-based MJCF is exported and re-imported. This breaks expectations for deterministic ordering in compiled models and downstream control/validation.

wp.tile_load_indexed reads out of bounds for negative indices, even though OOB indices are intended to predicate to zero. The upper bound is handled but negative indices slip through and dereference before src.data.

Warp under-allocates shared memory for a custom backward kernel when the backward requires more shared memory than the forward. This leads to out-of-bounds shared memory writes and CUDA error 700, while CPU appears to tolerate it.

MuJoCo’s USD decoder does not preserve several semantics needed for roundtrip validation: sites, model names, unlimited joints, and disabled-collider metadata. This undermines compiled-model comparisons and deterministic rollouts after import.

Mesh SDF generation settings are partially configurable and inconsistent across USD/URDF/MJCF imports. Request is to expose configuration paths for Mesh.build_sdf() settings and avoid import paths that drop sdf_* options before expensive cooks occur.

Automatic USD mesh approximation calls ModelBuilder.approximate_meshes() without forwarding method-specific settings. This blocks users from applying required settings (e.g., CoACD threshold migration) through ModelBuilder.add_usd() while respecting USD-authored approximation routing.

Inverse dynamics tests are intermittently failing in the merge queue and are reported as flaky. The failures need root-cause analysis to stabilize CI coverage.

Request to warn when CollisionPipeline auto-sizes a very large rigid contact buffer allocation. The warning should include the resolved slot count/bytes, drivers (pair count, world count), and remediation (set rigid_contact_max explicitly).

In Isaac Lab camera output example, point cloud generation fails because camera pose fields become zero/NaN. The report uses Isaac Lab main with Isaac Sim 5.1 and the documented camera output tutorial script.

With Isaac Sim 6.0.0 and Isaac Lab manager-based envs using `use_fabric=True`, `reset()` can return camera images from before the reset even though state observations are post-reset. `num_rerenders_on_reset` appears to rerender old transforms, causing the first inference each episode to see the wrong scene.

Request to standardize Solver.reset world_mask shape across solvers to (world_count + 1,), reserving the final slot for global entities with world = -1. It proposes updating the public contract with deprecation where needed and making solvers that lack global entities treat the final slot as a consistent no-op.

Newton's import_usd.py has grown beyond 4500 lines and is requested to be split into multiple files. The goal is improved readability and maintainability of the USD import pipeline.