Skip to main content
The rfx.decorators module provides the @policy decorator for marking functions as deployable and MotorCommands for building actions from named joints. 
import rfx

@rfx.policy
def hold(obs):
    return rfx.MotorCommands({"gripper": 0.8}, config=rfx.SO101_CONFIG).to_tensor()

# Then:
#   rfx deploy my_policy.py --robot so101

​
Decorator

​
@policy

Mark a function as an rfx policy for deployment with rfx deploy. A policy is any callable: Dict[str, Tensor] -> Tensor. This decorator stamps _rfx_policy = True so that rfx deploy my_file.py can discover it automatically. 
import rfx
import torch

@rfx.policy
def my_policy(obs):
    return model(obs["state"])

# Or with parentheses
@rfx.policy()
def my_policy(obs):
    return model(obs["state"])
​
fn
Callable | None
The function to decorate (when used without parentheses).
Returns: Callable - The decorated function, unchanged except for the _rfx_policy marker.

​
Classes

​
MotorCommands

Build action tensors from named joint positions. Instead of writing action[0, 5] = 0.8 and hoping index 5 is the gripper, use joint names from the robot config: 
from rfx import MotorCommands, SO101_CONFIG

cmd = MotorCommands(
    {"gripper": 0.8, "elbow": -0.3},
    config=SO101_CONFIG
)
action = cmd.to_tensor()  # shape: (1, max_action_dim)
Works with any robot — resolves joint names from the config’s joint list.
​
positions
dict[str, float] | None
Dict mapping joint name to target position.
​
config
RobotConfig | None
RobotConfig with joints list. Required for to_tensor().
​
kp
float
default:"20.0"
Position gain (for backends that use it).
​
kd
float
default:"0.5"
Damping gain (for backends that use it).

​
Methods

to_tensor(batch_size=1) Convert named positions to a padded action tensor. Returns a tensor of shape (batch_size, max_action_dim) with joint positions placed at the correct indices from the robot config. 
cmd = MotorCommands({"gripper": 0.8}, config=SO101_CONFIG)
action = cmd.to_tensor(batch_size=1)  # (1, max_action_dim)
​
batch_size
int
default:"1"
Leading batch dimension.
Returns: torch.Tensor - Action tensor ready to pass to robot.act(). Raises: ValueError - If config is not set or a joint name is not found.
to_list() Convert to a flat position list using config joint ordering. Returns a list of length action_dim with positions at the correct indices. Joints not specified default to 0.0. 
cmd = MotorCommands({"gripper": 0.8}, config=SO101_CONFIG)
positions = cmd.to_list()  # [0.0, 0.0, ..., 0.8, ...]
Returns: list[float] - Flat position list. Raises: ValueError - If config is not set.
MotorCommands.from_positions(positions, config=None, kp=20.0, kd=0.5) (classmethod) Create commands from named positions. 
cmd = MotorCommands.from_positions(
    {"gripper": 0.8, "elbow": -0.3},
    config=SO101_CONFIG,
    kp=30.0
)
​
positions
dict[str, float]
required
Dict mapping joint name to target position.
​
config
RobotConfig | None
RobotConfig with joints list.
​
kp
float
default:"20.0"
Position gain.
​
kd
float
default:"0.5"
Damping gain.
Returns: MotorCommands - New MotorCommands instance.

​
Example: Complete Deployable Policy

# my_walking_policy.py
import rfx
import torch

# Load a trained model
loaded = rfx.load_policy("hf://rfx-community/go2-walk-v1")

@rfx.policy
def walk(obs):
    """Walk forward at constant velocity."""
    # Use the loaded policy
    return loaded(obs)

# Deploy with:
#   rfx deploy my_walking_policy.py --robot go2

# my_manual_control.py
import rfx
from rfx import MotorCommands, GO2_CONFIG

@rfx.policy
def stand(obs):
    """Hold standing pose."""
    positions = {
        "FL_hip": 0.0,
        "FL_thigh": 0.9,
        "FL_calf": -1.8,
        "FR_hip": 0.0,
        "FR_thigh": 0.9,
        "FR_calf": -1.8,
        "RL_hip": 0.0,
        "RL_thigh": 0.9,
        "RL_calf": -1.8,
        "RR_hip": 0.0,
        "RR_thigh": 0.9,
        "RR_calf": -1.8,
    }
    return MotorCommands(positions, config=GO2_CONFIG).to_tensor()

# Deploy with:
#   rfx deploy my_manual_control.py --robot go2

Build docs developers (and LLMs) love

Get started for freeTalk to us