isaac_ros_foundationpose 5.0.0: no pose output on RTX PRO 3000 Blackwell Laptop GPU (sm_120), stuck with "Processing timeout"

Abierto
  • ¿Algo no está bien?
#103 0 comentarios 0 reacciones 0 asignados Ver en GitHub

Nadie ha tomado este issue todavía.

Evaluación

Dificultad
4/5
Tiempo estimado
3-5 días
Aptitud para principiantes
48/100
Tipo de issue
Error
Claridad
Bastante claro
Estado de actividad
Activo
Área
machine-learning

Línea de trabajo

Start by checking the FoundationPose node and CV-CUDA compatibility for sm_120; the report identifies libfoundationpose_node.so, libcvcuda.so.0.14.0, and isaac_ros_foundationpose_core.launch.py. Reproduce with the FoundationPose rosbag quickstart on the listed GPU and inspect which pipeline stage stops progressing. Done means pose output is produced or the unsupported stage and a workable path are identified.

Escrito por el modelo de indexación a partir del texto del issue.

Descripción

Hello,

I am running the FoundationPose rosbag quickstart (Mustard) and FoundationPose never publishes a pose on a Blackwell laptop GPU. Everything before FoundationPose works.

System
  • GPU: NVIDIA RTX PRO 3000 Blackwell Generation Laptop GPU, 12 GB, compute capability 12.0
  • Driver: 595.99.02
  • Host: Ubuntu 24.04, x86_64
  • Isaac ROS 5.0 (Docker, isaac-ros-cli 2.6.0), ROS 2 Lyrical
  • ros-lyrical-isaac-ros-foundationpose 5.0.0, TensorRT 10.16.1, CUDA 13.2 (container), libcvcuda0 0.14.0
  • Engines rebuilt inside the 5.0 container with install_foundationpose_models.sh and install_rtdetr_models.sh
Command

Official quickstart: ros2 launch isaac_ros_examples isaac_ros_examples.launch.py launch_fragments:=foundationpose ... (Mustard mesh, quickstart_interface_specs.json), then ros2 bag play -l quickstart.bag.

What happens
  • All engines load, [FoundationPoseNode] Initialization complete
  • /image_rect 8 Hz, /detections_output 0.56 Hz, /segmentation 0.55 Hz → RT-DETR and the mask work
  • /pose_estimation/output: nothing, ever
  • Log: [FoundationPoseNode] Processing timeout, releasing gate every 5–15 s
  • GPU stays at 100 % utilization, ~6.9 GB used, no crash
  • With pose_estimation_timeout_ms: 120000 (added in isaac_ros_foundationpose_core.launch.py), still no pose after 120 s → one frame never finishes
Observations
  • libfoundationpose_node.so 5.0.0 contains sm_120 SASS. libcvcuda.so.0.14.0 only has sm_70–sm_89 SASS (PTX sm_70) and is the only GPU library loaded in the container without sm_120 (checked with cuobjdump --list-elf on all libraries mapped by component_container_mt).
  • With Isaac ROS 4.6 (Jazzy), libfoundationpose_node.so only had sm_75. The node crashed with NVCV_ERROR_OUT_OF_MEMORY (11.7 GB peak), and once with RasterizeCudaFwdShaderKernel launch failed ... cudaErrorIllegalAddress.
  • The original FoundationPose (Python, NVlabs GitHub) runs fine on the same GPU (~26 ms per tracking frame).
Questions
  1. Is Blackwell sm_120 (laptop RTX PRO / RTX 50) supported by isaac_ros_foundationpose 5.0?
  2. Is a CV-CUDA build with sm_120 planned, or is there a workaround?
  3. Is there a debug option to see which stage of the FoundationPose pipeline is stuck?

Thank you!

Lenguaje dominante
C++
Estrellas
515
Forks
57
Sin PR fusionados en 30 d

Preparar el entorno

Primeros pasos

  1. Lee el issue completo y luego la guía de contribución del proyecto.
  2. Comenta en el issue que vas a ocuparte — evita que dos personas hagan lo mismo.
  3. Haz un fork del repositorio y trabaja en una rama.
  4. Abre un pull request que haga referencia al número del issue.

Más de NVIDIA-ISAAC-ROS/isaac_ros_pose_estimation

Todos los issues de NVIDIA-ISAAC-ROS/isaac_ros_pose_estimation

Issues similares

Más issues de C++

Recibe los nuevos issues en tu correo

Un resumen breve de issues de GitHub para principiantes.