Deploy API - Reflex

rfx.deploy() is the main entry point. It handles everything: load policy, resolve robot config, create robot, run the control loop.

Signature

stats = rfx.deploy(
    "runs/my-policy",       # path, hf:// URL, or .py file
    robot="so101",           # robot type or YAML path
    port="/dev/ttyACM0",     # optional hardware override
    rate_hz=50,              # control frequency
    duration=30,             # seconds (None = infinite)
    mock=False,              # use MockRobot instead
    device="cpu",            # torch device
)

Parameters

Parameter	Type	Default	Description
`policy`	`str \| Path`	required	Path to policy dir, `hf://org/repo`, or `.py` file.
`robot`	`str \| Path`	auto	Robot type (`so101`, `go2`, `g1`) or YAML config path.
`port`	`str`	auto	Serial port or IP override.
`rate_hz`	`float`	config	Control loop frequency. Defaults to robot config.
`duration`	`float \| None`	`None`	Run time in seconds. `None` runs until Ctrl+C.
`mock`	`bool`	`False`	Use `MockRobot` instead of real hardware.
`device`	`str`	`"cpu"`	Torch device for inference.
`warmup`	`float`	`0.5`	Seconds to sleep after reset before starting.

Return value

Returns a stats object with timing information:

stats.iterations      # total control loop steps
stats.overruns        # steps that missed their deadline
stats.p50_jitter_s    # median per-step jitter
stats.p95_jitter_s    # 95th percentile per-step jitter

Auto-detection

If the policy artifact stored a robot config (via policy.save(..., robot_config=config)), robot is inferred automatically. Explicit robot or config overrides the stored value.

Examples

rfx.deploy("runs/my-policy", robot="so101")

rfx.deploy("hf://user/policy", robot="go2", duration=30)

rfx.deploy("my_policy.py", robot="so101")

rfx.deploy("runs/my-policy", robot="so101", mock=True)