Search — cuda

Depending on SimulationApp startup/import order, Isaac Lab classes can be imported twice and end up with mismatched parent classes. This breaks type checks (e.g., config objects failing isinstance) and can cause subtle downstream failures.

Warp lowering of conditional merges can leave operands uninitialized along some paths, and a nested case leaked a branch value into its sibling branch. Under register pressure this surfaced as illegal CUDA memory access.

wp.tile_load_indexed performs an out-of-bounds read when given a negative index. Upper-bound OOB is predicated to zero, but negative indices can dereference before src.data instead of producing 0.

A previously disabled friction ramp test (mujoco_warp_cuda_0) needs follow-up to be re-enabled. The issue provides no additional details beyond the re-enable request.

On Isaac Lab’s Newton backend, a hard reset can crash with CUDA error 700 on the first step because CollisionPipeline reuses cached references to a freed/re-finalized Newton Model. Under GPU memory pressure, the next collide dereferences freed or mismatched device buffers.

On Windows with Isaac Sim 6.0.1 prebuilt binaries, the VS Code setup task generates python.analysis.extraPaths that conflicts with pyproject.toml in a newly created Isaac Lab project. This produces a VS Code/Pylance configuration conflict rather than a clean setup.

Changing Warp CPU compiler flags between modules in the same process triggers Clang target-feature errors and forces Warp to discard PCH and recompile without it. Tests pass but compilation becomes slower and logs misleading errors.

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.

Freeing a CUDA array while capturing a conditional body graph can add a mem-free node to the wrong CUDA graph. This triggers cudaGraphAddMemFreeNode failures (cudaErrorInvalidValue) and breaks capture workflows.

A ROS2 camera OmniGraph driven by OnPlaybackTick stops publishing after 1–2 frames when added as an `over` on a payload-referenced robot. Render products and image data appear valid, but ROS2CameraHelper output stalls.

Users request a design document covering API Capture, serialized .wrp graphs, and CPU capture/replay. The feature set has expanded enough that code and user guide usage examples are not sufficient to understand behavior and constraints.

Request to support block_dim > 1 on CPU with CUDA-like block semantics, including shared tile state and barriers. Proposal suggests implementing cooperative fibers per block to preserve correctness while improving efficiency.

Warp can retain deterministic launch metadata when switching to a cached non-deterministic kernel variant of the same module. This results in incorrect execution (e.g., counters run twice), and reproduces only with a warm disk cache.

User hits ERROR_OUT_OF_DEVICE_MEMORY on a GCP setup using an NVIDIA L4 while following the recommended Isaac Sim cloud setup guide. The issue occurs with both compiled-from-source and precompiled binaries, preventing progress.

The kamino_mujoco_admm_solver example consistently warns about truncated contacts. The log indicates this happens in a typical setup (Warp 1.15.0 on A40), suggesting defaults or limits are easy to exceed.

The `admm_contact_solver` example sometimes crashes when clicking the UI "Reset" button, with Warp CUDA error 700 illegal memory access. The crash is intermittent and can occur after several resets.

In Isaac Sim 6.0.0 headless (`--no-window`), rendering paths that call `app.update()` stall at ~10 seconds per frame after a viewport creates a surfaceless render product. The render fence never signals in headless mode, blocking RTX sync until an internal timeout.

ControllerPID.State.reset(mask) launches a kernel without specifying a device, so it may run on CPU even when inputs are on CUDA. This can cause a crash (exit code 139) instead of resetting masked entries, and should validate mask compatibility up front.

ArticulationView accepts invalid boolean masks without validating shape/device, and a subsequent kernel may read out of bounds. An empty CUDA mask can trigger CUDA illegal memory access (error 700) and poison the CUDA context.

ModelBuilder.add_particles() allows inconsistent array lengths across positions/velocities/masses/etc., producing an invalid model that still finalizes. This can yield non-finite state on CPU and potential out-of-bounds reads on CUDA; inputs should be validated before mutation and at finalize().

Examples that adjust ViewerGL camera speed fail because _cam_speed moved from ViewerGL to ViewerGUI in Newton v1.3.0. The example uses hasattr checks that no longer find the attribute, making small-scale scenes hard to view.

wp.quat_twist_angle() uses acos on the twist scalar component, which rounds to 1.0 in float32 near zero, producing a dead zone and discontinuity. Request is to stabilize behavior near zero and add a signed variant.

APIC capture of a padded bsr_set_transpose intermittently triggers heap corruption in CI (glibc free invalid next size) and AddressSanitizer reports a heap-buffer-overflow. The crash breaks the process pool and fails the job.

Newton reconstructs signed revolute joint coordinates using acos(twist[3]) which becomes zero in float32 for small rotations, creating a dead zone and discontinuity. The signed information exists in the quaternion vector but is lost by the acos path.

Switching examples after enabling certain visualization options (Show Joints/Contacts/Inertia Boxes) can crash with Warp CUDA error 700, leading to PTX module load failures. This occurs when moving from robot_g1 to admm_contact_solver.

With the widespread availability of parallel computing hardware, sampling-based motion planning methods such as Model Predictive Path Integral (MPPI) control have become increasingly powerful for complex nonlinear systems in non-smooth task spaces. However, the sampling and forward-simulation pipeline in MPPI suffers from averaging-induced failure in cluttered environments, where the importance-weighted update averages incompatible rollouts and leads to hesitation or even collision when an obstacle lies directly ahead. This paper proposes Clustering-Embedded MPPI (CE-MPPI), a framework that architecturally resolves the averaging-induced failures inherent in standard MPPI within non-convex environments. Rather than simply mitigating interference, CE-MPPI redefines the control law by integrating a high-fidelity pruning and clustering stage. By leveraging density-based spatial clustering of applications with noise (DBSCAN) alongside a novel geometric direction feature that is extracted from collision-derived reference points, the system isolates feasible trajectory modes from the noise of infeasible rollouts. This is paired with an intelligent selection logic that optimizes for minimum cost in static scenes while actively steering opposite to obstacle flux in dynamic environments. Experiments in 2-D JAX-accelerated simulations show that CE-MPPI alleviates obstacle-front hesitation and avoids persistent coupling with moving obstacles in dynamic scenes. In particular, real-world tests on a 6-DoF UR5e manipulator with CUDA-parallel rollouts in Isaac Gym achieve a 48\% reduction in time-to-goal and a 12\% shorter end-effector path.