Search — warp

Warp’s Quick Start example fails unless the user makes a manual code change, throwing a codegen overload/typing error for extract(). This blocks new users from running the introductory workflow.

Warp’s experimental cuBQL wp.Mesh backend only supports ray queries; point, furthest-point, and AABB query APIs error under debug builds or return empty results. The request is to extend cuBQL traversal support so it becomes a drop-in replacement backend.

Some Warp built-ins are marked differentiable even though their native adjoints are stubs or incorrect, so backward passes compile and run but produce wrong/zero gradients. Vector wp.sign() is called out as producing incorrect gradients for vectors.

newton.eval_fk lacks handling for JointType.CABLE, leaving transforms at identity and overwriting child body poses to rest offsets. The failure is silent and corrupts articulation state for cable-joint mechanisms.

Upgrading Newton from v1.0.0 to v1.2.0 reportedly reduces hydroelastic contact forces ~30× for the same parameters in a wheel-inside-pipe case. This is described as a breaking change for tuned controllers and suggests inconsistent calibration across releases.

A GitHub issue titled 'PEAK' contains unrelated text and no actionable Warp request. This appears to be spam or off-topic content.

Warp may retain discarded temporary modules when repeatedly declaring equivalent `@wp.kernel(module="unique")` kernels. This can keep Python object graphs reachable and grow memory over repeated factory calls.

A dexterous hand imported via URDF cannot grasp a bottle reliably; the bottle slides and cannot be lifted. The reporter notes the Franka example can lift the same object, implying a hand-specific contact/friction issue.

Warp plans work to support GPU kernel launches with hardware-coherent CPU memory, pinned CPU arrays, and peer GPU arrays when directly addressable. The issue emphasizes preserving clear diagnostics for invalid cases.

Newton's scheduled nightly workflow failed specifically in the Warp nightly tests suite while other suites passed. The issue references the failing GitHub Actions runs and logs.

Proposal to integrate DiffRL into IsaacLab via an isaaclab_diffrl extension and the Mineral algorithm library. It targets the Direct workflow and Newton (Warp) backend to enable end-to-end backprop through physics for algorithms like SHAC and APG/BPTT.

Running the full Newton test suite occasionally triggers out-of-bounds errors, seen in some CUDA example tests. It is hard to reproduce in isolation, suggesting an order- or resource-dependent issue.

Warp docs expose dtype-specific HashGrid query types alongside the intended wp.HashGridQuery, which users shouldn't need to handle. The request is to consolidate documentation/stubs around the single public type while keeping internal codegen behavior.

Some array-rooted composite augmented assignments and matrix row writes either fail codegen or silently drop gradients on current Warp main. This breaks differentiable programs that rely on these write patterns.