Actuators and sensors

For gravity-gradient, atmospheric drag, solar pressure, and residual magnetic loads, see forces and disturbances.

Standard MuJoCo actuator controls live in data.ctrl. Spacecraft devices use separate buffers in data.actuators, populated in the order the devices were added to the model specification.

Device	Runtime command	Units	Limits
Reaction wheel	rw_torque_cmd	N·m	Optional torque and wheel-speed limits
Magnetorquer	mtq_dipole_cmd	A·m²	Symmetric dipole limit; magnetic model must be enabled
Thruster	thr_force_cmd	N	Clamped from zero to force_limit
Control moment gyro	cmg_gimbal_rate_cmd	rad/s	Optional gimbal-rate and angle limits

Commands are held across steps until overwritten or reset. Reaction-wheel torque changes rotor momentum and applies the opposite torque to the attached body. Thruster direction and application point are body-frame metadata. CMGs are currently available only on the CPU backend.

After building the reaction-wheel model from modeling, create data with the orbit from the quick start, then command it:

data.actuators.rw_torque_cmd[0] = 0.01  # N·m
mjo_step(model, data)

Only index a device buffer when that device exists in the compiled model. A free-body model with no wheel has an empty rw_torque_cmd array.

Sensors

Declare sensors in the MuJoCo XML. model.sensor(name) and model.sensors.descriptors expose metadata, while data.sensordata contains the native sensor values. The CPU measurement namespace adds bias/noise handling:

measurements = data.sensors.measure_all(noisy=False)
noisy_measurements = data.sensors.measure_all(noisy=True)

Pass rng_seed to model.make_data(...) for reproducible runtime noise. For one configured sensor, use data.sensors.measure(name, noisy=True). Axis and quaternion measurements use rotational perturbations and remain normalized. The GPU backend exposes native sensor buffers but does not yet implement the noisy measurement namespace.

MuJoCo evaluates sensors before integrating a step. Rollout sensor row k therefore corresponds to the start of step k, while packed state row k is the state after the step. Account for this offset when mixing states and sensors in a cost or comparison.