# FoundationPose local setup

Validated 2026-09-16 for model-based RGB-D inference using an existing STL.

- Source: `third_party/FoundationPose`, upstream commit `a1b694b83e633c2cb6115b9063d940a687759392`.
- Interpreter: existing `.venv/bin/python`; reused Torch 2.7.1+cu128 and PyTorch3D.
- CUDA toolkit: existing `.cuda/usr/local/cuda-12.8`; GPU architecture 8.6.
- Installed upstream Python requirements, pybind11, ninja, and locally built nvdiffrast 0.4.0.
- Compiled `third_party/FoundationPose/mycpp/build/mycpp.cpython-311-x86_64-linux-gnu.so`.
- Weights and mirror provenance: `third_party/FoundationPose/weights/download_manifest.json`.

## Run the verified smoke check

From the HandFlow root, in a shell with NVIDIA GPU access:

```bash
source scripts/foundationpose_env.sh
python scripts/verify_foundationpose_setup.py
```

The environment script clears ROS-related Python/library overrides and selects the existing project environment/toolkit. It does not install Torch or CUDA.

The check exercises actual CUDA rasterization, loads both pretrained networks and executes their forward passes, and initializes FoundationPose with `docs/上半.stl`, including compiled C++ rotation clustering. Results: `output/foundationpose_setup/runtime_verification.json`; installation/build logs are in the same directory.

This does not validate pose accuracy or contact constraints on the recorded video. Optional BundleSDF `mycuda`/NeRF extensions are not required for the existing-STL path and were not built.

## Rebuild compiled modules

```bash
source scripts/foundationpose_env.sh
python -m pip install --no-build-isolation --no-deps ./third_party/nvdiffrast
cmake -S third_party/FoundationPose/mycpp -B third_party/FoundationPose/mycpp/build \
  -DCMAKE_BUILD_TYPE=Release \
  -Dpybind11_DIR="$FP_ROOT/.venv/lib/python3.11/site-packages/pybind11/share/cmake/pybind11" \
  -DPYTHON_EXECUTABLE="$FP_ROOT/.venv/bin/python" \
  -DPython_EXECUTABLE="$FP_ROOT/.venv/bin/python"
cmake --build third_party/FoundationPose/mycpp/build -j2
```