Daily Digest

Loading brief…

• Kernel requests too little shared memory when it moves a tile returned by a `wp.func` to shared storage

  github:NVIDIA/warp11/11/2026crashes-stabilitywarpcudashared-memorytilekernelillegal-memory-access

  Warp CUDA kernels that call a wp.func returning a tile can under-request dynamic shared memory when the tile is later moved to shared storage (e.g., via wp.tile_broadcast). This can silently corrupt results or trigger illegal memory accesses depending on other

  ⚡ Correct CUDA shared-memory sizing in Warp kernels is critical for trustworthy high-performance simulation and differentiable workloads in Physical AI pipelines.

• [BUG] Actuator gradients are wrong with several steps on one tape

  github:newton-physics/newton11/11/2026crashes-stabilitynewton-physicswarpautodiffgradientstapeactuator

  Newton Actuator steps reuse internal buffers, so recording multiple steps on one wp.Tape causes later steps to overwrite data needed for earlier backward passes. This produces incorrect gradients for multi-step tapes, especially when nonlinear components are i

  ⚡ Correct differentiable physics gradients are essential for training and optimization loops that underpin Physical AI control and design workflows.

• Custom AGX Orin Carrier board stuck at UEFI

  nvidia-forum:robotics-edge-computing11/11/2026hardware-integrationjetson-agx-orincarrier-boarduefibootbringuphardware-debug

  A custom Jetson AGX Orin carrier board is stuck at UEFI. This prevents booting into the OS and blocks any further software setup or deployment work.

  ⚡ Robust bring-up on custom carrier boards is key to making NVIDIA edge platforms viable for real-robot deployments.

• [BUG] wp.Mesh crashes on out-of-range triangle indices

  github:NVIDIA/warp11/10/2026crashes-stabilitywarpmeshsegfaultbounds-checkinggeometrycpu

  Warp wp.Mesh() does not bounds-check triangle indices and can segfault on CPU with out-of-range or negative indices. Mesh.refit() and assigning Mesh.points can also crash because they read through indices without validation.

  ⚡ Hard crashes from invalid geometry inputs erode developer confidence and complicate productionizing simulation and geometry pipelines.

• [REQ] Make wp.sparse.bsr_mm differentiable (tape.backward currently crashes)

  github:NVIDIA/warp11/10/2026training-infrawarpsparsebsr-mmautodifftapecrash

  Recording warp.sparse.bsr_mm on a Tape causes tape.backward() to crash due to indexedarray lacking .grad. Even without the crash, kernels are compiled with enable_backward=False, so gradients would be missing.

  ⚡ Differentiable sparse linear algebra is foundational for scalable optimization and learning pipelines that rely on structured physics and constraints.

• Hardware timestamping on RTL8126 fails

  nvidia-forum:robotics-edge-computing11/11/2026hardware-integrationjetsonnicrtl8126hardware-timestampingptpedge-computing

  Hardware timestamping fails on the RTL8126 network interface. This blocks or degrades time-synchronized networking workflows that depend on NIC timestamp support.

  ⚡ Reliable hardware timestamping is foundational for synchronized sensing and control loops across edge robotics deployments.

• Jetson Thor USB Host Controller Resource Limit with Multiple CAN Adapters

  nvidia-forum:robotics-edge-computing11/11/2026hardware-integrationjetson-thorusbhost-controllercan-buscan-adapterresource-limit

  Jetson Thor USB host controller hits a resource limit when using multiple CAN adapters. This prevents scaling to multiple CAN devices on the same system without changes or workarounds.

  ⚡ I/O scalability limits on edge platforms directly constrain real-robot sensor/actuator integration for Physical AI applications.

• HSB SDK v3.0.0: How to access data in a custom Python operator?

  nvidia-forum:robotics-edge-computing11/11/2026docs-onboardinghsb-sdkpython-operatorcustom-operatordata-accesssdk-usagerobotics-edge

  A user asks how to access data in a custom Python operator in HSB SDK v3.0.0. This suggests missing or unclear documentation/API patterns for operator I/O and data handling.

  ⚡ Better operator data-access guidance reduces friction when extending NVIDIA edge/robotics SDK pipelines used in Physical AI products.

• multicontact: GJK/MPR manifold scratch buffers rely on CUDA-only fixedarray idioms

  github:newton-physics/newton11/11/2026integrationnewton-physicswarpmulticontactcuda-specificmetal-backendportability

  Newton multicontact manifold scratch buffers rely on CUDA-only fixedarray and pointer aliasing behaviors. This breaks on backends with strict address spaces (e.g., Apple Metal) because it treats thread-local stack storage as a device pointer.

  ⚡ Reducing backend-specific assumptions in core physics/contact code increases portability across GPU backends needed by diverse Physical AI developer environments.

• Nvidia,disable-power-down

  nvidia-forum:robotics-edge-computing11/11/2026hardware-integrationjetsonpower-managementdevice-treedisable-power-downplatform-configedge-computing

  The topic references "Nvidia,disable-power-down", indicating a need to disable a power-down behavior via NVIDIA-related configuration. Missing context suggests unclear device-tree or platform settings for power management.

  ⚡ Power-management controls that are hard to configure can destabilize real-time robotics behavior and complicate deployment on NVIDIA edge devices.

• How can we run inference of Alpamayo 2 super on jetson AGX thor?

  nvidia-forum:robotics-edge-computing11/11/2026deploymentjetson-agx-thorinferencedeploymentmodel-compatibilityedge-airuntime

  A user asks how to run inference for "Alpamayo 2 super" on Jetson AGX Thor. This indicates uncertainty about model support, optimization, or runtime deployment steps on that platform.

  ⚡ Running modern model inference reliably on Jetson-class hardware is central to deploying Physical AI capabilities on robots.

• <model> asset in an included file is looked up with the directory twice when the main file is loaded as dir/scene.xml

  github:google-deepmind/mujoco11/10/2026asset-pipelinemujocomjcfasset-resolutionincludepath-bugrobot-models

  MuJoCo incorrectly resolves a <model> asset file path in an included file when the main MJCF is loaded from a subdirectory, duplicating the directory in the resolved path. This causes model loading to fail with a file-not-found error like 'case2/case2/sub/chil

  ⚡ Consistent asset path resolution is important for reliable robot model composition workflows that feed simulation-to-training pipelines.

• to_xml() drops mass="0" density="0" from a geom default class, so the saved XML compiles to a heavier body

  github:google-deepmind/mujoco11/10/2026asset-pipelinemujocomjcfxml-exportmodel-serializationmass-propertiesrobot-arms

  MuJoCo's to_xml() drops explicit mass="0" and density="0" on geom defaults, changing the exported model's physical properties. The re-saved XML compiles with default density and yields a much heavier body than the original.

  ⚡ Simulation-to-asset export fidelity issues directly undermine trust in model interchange and sim2real validation workflows.

• Support wp.handle in framework and DLPack converters

  github:NVIDIA/warp11/10/2026integrationwarpdlpackjaxpytorchpaddlehandle-dtype

  Warp converters (from_torch/from_jax/from_paddle/from_dlpack) do not support explicit dtype=wp.handle imports/exports using 64-bit integer storage. Users want handle arrays to round-trip through DLPack and framework converters while preserving underlying bits

  ⚡ Robust interop (Torch/JAX/Paddle/DLPack) is a key enabler for Physical AI pipelines that mix simulation, learning, and GPU compute libraries.

• wp.sqrt() gradient is inf at x == 0, unlike wp.length()

  github:NVIDIA/warp11/10/2026training-infrawarpautodiffgradientsnumerical-stabilitysqrtnan

  Warp's wp.sqrt() adjoint computes 0.5/sqrt(x) with no guard, producing inf gradients at x==0 that often become NaNs downstream. This differs from wp.length() behavior and can break common patterns like sqrt(dot(v,v)) at v==0.

  ⚡ Numerical stability in autodiff primitives is critical for reliable gradient-based optimization and learning workloads in Physical AI.

• [REQ] Make wp.sparse.bsr_mv differentiable

  github:NVIDIA/warp11/10/2026training-infrawarpsparsebsr-mvautodifftapegradients

  warp.sparse.bsr_mv is not differentiable with wp.Tape because kernels are built with enable_backward=False. tape.backward() completes but returns all-zero gradients for x and A.values with only a warning, requiring users to implement their own adjoint.

  ⚡ Missing gradients in sparse ops force users into custom adjoints, slowing adoption of Warp as a learning/optimization backend for Physical AI.

• [Python] rollout rejects a tuple of MjData although its signature accepts a Sequence

  github:google-deepmind/mujoco11/10/2026tooling-dxmujocopython-apirollouttypingsequencewindows

  MuJoCo's Python rollout API claims to accept a Sequence[MjData] but only recognizes list, so tuples get wrapped incorrectly and then fail in compiled code. The resulting error message is unhelpful and doesn't explain the real cause.

  ⚡ API contract mismatches in core rollout tooling increase user confusion and slow iteration in simulation-driven learning loops.

• 30/5000 Hello, NVIDIA team. How does Yocto support the overlays function?

  nvidia-forum:robotics-edge-computing11/10/2026docs-onboardingnvidia-forumsyoctodevice-tree-overlaysrobotics-edgeembedded-linuxos-integration

  A forum user asks how Yocto supports the overlays function, with no additional context provided. The lack of details suggests a documentation or onboarding gap around overlays in NVIDIA's Yocto-related support.

  ⚡ Clear guidance for embedded OS customization (like Yocto overlays) reduces integration risk for edge robotics deployments.

• NVIDIA Jetson Thor T5000 module (699-pin, 128GB LPDDR5X)

  nvidia-forum:robotics-edge-computing11/11/2026hardware-integrationjetson-thort5000modulelpddr5xcarrier-boardhardware-specs

  The post asks about the NVIDIA Jetson Thor T5000 module with a 699-pin interface and 128GB LPDDR5X. It indicates interest in hardware specs/availability details for integration planning.

  ⚡ Clear, accessible module configuration information helps developers plan edge compute deployments for Physical AI workloads.

• Support union type annotations on kernel and function parameters

  github:NVIDIA/warp11/11/2026tooling-dxwarptype-annotationsunion-typeskernel-apiindexedarraypython-typing

  Warp users want union type annotations on kernel/function parameters so a parameter can accept a fixed set of types (e.g., wp.array or wp.indexedarray). This would enable writing one kernel that works on full arrays or indexed subsets without separate overload

  ⚡ Stronger typing ergonomics in Warp improves developer velocity for writing reusable kernels used across simulation and learning stacks.

• Share rod and cable kernels across solvers

  github:newton-physics/newton11/10/2026tooling-dxnewton-physicssolver-kernelsrodscablescode-organizationvbd

  Newton has duplicated rod/cable energy kernel logic across solvers (VBD and Kamino). The request is to move rod kernels into a shared module to preserve behavior while aligning implementations.

  ⚡ Shared physics kernels reduce duplication and help keep solver behavior consistent across modules used in robotics simulation stacks.

• Add wp.no_grad(), Tape.get_active() and nested tapes

  github:NVIDIA/warp11/10/2026tooling-dxwarpautodifftapeno-gradapi-designdeveloper-experience

  Warp lacks public APIs to pause/query the active Tape and to support nested tapes without reaching into internal runtime state. The request proposes wp.no_grad(), wp.Tape.get_active(), and nested tape support to control recording safely.

  ⚡ First-class gradient recording controls enable cleaner library integrations and reduce reliance on private internals in Physical AI compute stacks.