[ARM64] TensorRT engine build failed for autoware_bevfusion and autoware_lidar_transfusion on NVIDIA AGX Orin (JetPack 6.2)

[cccdotvvv]

Checklist

• I've read the contribution guidelines.
• I've searched other issues and no duplicate issues were found.

Description

Environment / Setup

• Docker Image: autoware-devel-20260402-cuda-arm
• Host Machine: NVIDIA AGX Orin Developer Kit
• JetPack Version: 6.2
• Autoware Code Version: 1.7.1

Docker run command:

bash

docker run -it \
  --runtime=nvidia \
  --gpus all \
  --privileged \
  --name autoware-test \
  --net=host \
  --shm-size=4gb \
  --env DISPLAY=$DISPLAY \
  --env NVIDIA_VISIBLE_DEVICES=all \
  --env NVIDIA_DRIVER_CAPABILITIES=all \
  --volume /tmp/.X11-unix:/tmp/.X11-unix:rw \
  --volume /work/workspace:/workspace \
  --device /dev/snd \
  ca7653815e6d \
  /bin/bash

Issue 1: autoware_bevfusion TensorRT Engine Build Failure

Command executed:

bash

ros2 launch autoware_bevfusion bevfusion.launch.xml build_only:=true

Error output:

text

[autoware_bevfusion_node-1] [E] [TRT] Error Code: 9: Skipping tactic 0x0000000000020306 due to exception Cask Gemm execution
[autoware_bevfusion_node-1] [E] [TRT] Error Code: 9: Skipping tactic 0x00000000000202d1 due to exception Cask Gemm execution
[autoware_bevfusion_node-1] [E] [TRT] Error Code: 9: Skipping tactic 0x000000000002000b due to exception Cask Gemm execution
[autoware_bevfusion_node-1] [E] [TRT] Error Code: 9: Skipping tactic 0x000000000002001d due to exception Cask Gemm execution
[autoware_bevfusion_node-1] [E] [TRT] Error Code: 9: Skipping tactic 0x00000000000201b0 due to exception Cask Gemm execution
[autoware_bevfusion_node-1] [E] [TRT] Error Code: 9: Skipping tactic 0x0000000000020348 due to exception Cask Gemm execution
[autoware_bevfusion_node-1] [E] [TRT] IBuilder::buildSerializedNetwork: Error Code 10: Internal Error (Could not find any implementation for node /bbox_head/decoder.0/cross_posembed/position_embedding_head/position_embedding_head.0/Conv + /bbox_head/decoder.0/cross_posembed/position_embedding_head/position_embedding_head.2/Relu.)
[autoware_bevfusion_node-1] [E] [TRT] [checkMacros.cpp::catchCudaError::212] Error Code 1: Cuda Runtime (no kernel image is available for execution on the device)
[autoware_bevfusion_node-1] [E] [TRT] Fail to create host memory
[autoware_bevfusion_node-1] [I] [TRT] Engine build completed
[autoware_bevfusion_node-1] terminate called after throwing an instance of 'std::runtime_error'
[autoware_bevfusion_node-1]   what():  Failed to setup TRT engine.

Issue 2: autoware_lidar_transfusion Node Crash During Engine Build

Command executed:

bash

ros2 launch autoware_lidar_transfusion lidar_transfusion.launch.xml build_only:=true

Error output:

text

[autoware_lidar_transfusion_node-1] [I] [TRT] Applying optimizations and building TensorRT CUDA engine. Please wait for a few minutes...
[autoware_lidar_transfusion_node-1] [I] [TRT] Applying optimizations and building TensorRT CUDA engine. Please wait for a few minutes...
[autoware_lidar_transfusion_node-1] [I] [TRT] Applying optimizations and building TensorRT CUDA engine. Please wait for a few minutes...
[autoware_lidar_transfusion_node-1] [I] [TRT] Applying optimizations and building TensorRT CUDA engine. Please wait for a few minutes...
[autoware_lidar_transfusion_node-1] [I] [TRT] Applying optimizations and building TensorRT CUDA engine. Please wait for a few minutes...
[ERROR] [autoware_lidar_transfusion_node-1]: process has died [pid 96, exit code -11, cmd '/opt/autoware/lib/autoware_lidar_transfusion/autoware_lidar_transfusion_node --ros-args --log-level info --ros-args -r __node:=lidar_transfusion --params-file /tmp/launch_params_z54p_2yg --params-file /tmp/launch_params_nf7sc3np --params-file /opt/autoware/share/autoware_lidar_transfusion/config/detection_class_remapper.param.yaml --params-file /opt/autoware/share/autoware_lidar_transfusion/config/transfusion_common.param.yaml --params-file /tmp/launch_params_jikrki_o -r ~/input/pointcloud:=/sensing/lidar/pointcloud -r ~/input/pointcloud/cuda:=/sensing/lidar/pointcloud/cuda -r ~/output/objects:=objects'].

Additional notes:

• Exit code -11 (SIGSEGV) suggests a segmentation fault during engine build.

Question

Are there any known compatibility issues with the combination of:

• Autoware 1.7.1
• JetPack 6.2 on AGX Orin
• The autoware-devel-20260402-cuda-arm Docker image

Specifically:

1. Is there a TensorRT version mismatch between what the image expects and what JetPack 6.2 provides?
2. Does autoware_bevfusion / autoware_lidar_transfusion require specific compilation flags for ARM64/Orin GPU architecture?
3. Are there any workarounds to successfully build the TensorRT engines for these modules on this platform?

Any guidance would be greatly appreciated. Thank you!