Bolt Runtime
Bolt any policy onto any robot. Bolt Runtime (boltrt) is a native C++
deployment runtime for learned robot policies.
- Modular observations: adapt a new policy by composing registered observation terms in its YAML.
- Seamless sim-to-real: MuJoCo and real robots implement one interface, decoupled from the policy, so switching is just
--backend. - Fast: native C++ runs the full policy step 3.25× faster than the Python stack.
A run is always the combination of three configurations plus an optional task:
boltrt [--task tracking] --robot g1 --policy policy.yaml --backend mujoco
| Piece | Owns |
|---|---|
| robot | joint/body names, MJCF model, dynamics and torque limits |
| policy | ONNX model, observations, PD gains, default pose, action scale |
| backend | where commands go: MuJoCo, Unitree G1, RoboParty RP1 or a recorded replay |
| task | what the loop does around the policy; tracking adds reference motion |
Where to go next
- New here: Installation, then Quick start.
- How the runtime is built: Core architecture.
- Running tracking policies, motions and teleop: Tracking task.
- Deploying on hardware: Real robots.
- Adding your own policy, robot or backend: Extending.
- Working with a coding agent: the repository ships a
boltrt-extendskill that walks agents such as Claude Code and Codex through these tasks.