Installation
Bolt Runtime builds on Linux x86_64 and aarch64 with CMake 3.22+ and a C++20 compiler. Python is not needed to build or run it.
Simulation build
sudo apt-get install cmake g++ curl libeigen3-dev libyaml-cpp-dev \
zlib1g-dev libzmq3-dev libglfw3-dev libgl1-mesa-dev
./tools/fetch-deps.sh --sim-only
cmake --preset sim && cmake --build --preset sim --parallel
ctest --preset sim
export PATH="$PWD/build/sim:$PATH"
fetch-deps.sh downloads pinned MuJoCo and ONNX Runtime releases (and the robot
SDKs when requested) into third_party/, verified against
dependencies.lock.json.
CMake presets and options
| Preset | Contents |
|---|---|
sim | MuJoCo backend with native viewer |
ci | sim without the viewer (headless machines) |
robots | sim plus the G1 and RP1 hardware backends |
| Option | Default | Effect |
|---|---|---|
BOLTRT_VIEWER | ON | Native MuJoCo viewer (needs OpenGL and GLFW) |
BOLTRT_G1 | OFF | Unitree G1 backend (real robots) |
BOLTRT_RP1 | OFF | RoboParty RP1 backend; links GPL-3.0 code (licensing) |
Robot assets
Robot configs reference models such as hf://elijahgalahad/g1_xmls@main/g1-mode_13_15.xml.
On first use Bolt Runtime downloads the whole repository snapshot into the
standard Hugging Face cache (~/.cache/huggingface/hub), so later runs and the
Python tools share it. It calls the system curl; no Python is involved.
| Variable | Effect |
|---|---|
HF_HUB_OFFLINE=1 | Never download; fail if the asset is not cached |
HF_ENDPOINT | Use a mirror instead of https://huggingface.co |
HF_TOKEN | Token for private repositories (also read from the HF token file) |
HF_HUB_CACHE / HF_HOME | Cache location |
On a robot without internet access, run once on a connected machine and copy
the cache, then set HF_HUB_OFFLINE=1.