Skip to main content

Backends

A backend is where robot state comes from and where joint commands go. Every backend implements the same four calls:

class RobotBackend {
public:
virtual void start() = 0;
virtual std::optional<RobotState> read_state() = 0; // nullopt: session closed
virtual void submit(const JointCommand &) = 0; // PD targets, gains, feedforward
virtual void stop() noexcept = 0;
};

The policy loop does not know which backend it runs on, so moving from simulation to hardware is a flag:

boltrt --robot g1 --policy policy.yaml --backend mujoco
boltrt --robot g1 --policy policy.yaml --backend real

MuJoCo​

type: mujoco
physics_dt: 0.005
realtime: false
floor: true
  • Online (default with a window): physics runs on its own thread at physics_dt using the latest command; the policy keeps its own clock.
  • Headless (--headless): no window and no pacing; physics and policy step deterministically, for batch evaluation and regression checks.
  • The native viewer owns its shortcuts (Space pause, Backspace reset). These affect physics only.

Hardware: G1 and RP1​

Hardware backends run a high-frequency command writer with a watchdog in the same process. If commands stop arriving within command_timeout_s, or state stops within state_timeout_s, the backend fails safe. Details, safety notes and configuration: Real robots.

Replay​

replay feeds recorded robot states into the loop without simulating anything, and discards the commands. Use it to reproduce a run offline, to debug a state sequence recorded on hardware, or for latency measurement and same-state parity checks.

# replay.yaml
type: replay
path: states.jsonl # e.g. the output of --record
boltrt --task tracking --robot g1 --policy policy.yaml --backend mujoco \
--motion walk.npz --headless --start --record states.jsonl
boltrt --task tracking --robot g1 --policy policy.yaml --backend replay.yaml \
--motion walk.npz --headless --start --record replayed.jsonl

Each line needs q, dq, orientation (wxyz), angular_velocity, tick and time; root_position is optional. Like MuJoCo, the replay advances to the next line only after a command is submitted, so tick i sees recorded line i and a replayed run reproduces the recorded actions. The run ends at the end of the file.

Adding a backend​

Implement RobotBackend, register a factory in BackendRegistry, and keep SDK headers and build rules inside src/backend/<name>/. See Add a backend.