mjlab (MuJoCo GPU)#
This page explains how to integrate BAM friction models into an mjlab
pipeline running on GPU with MuJoCo Warp. The entry point is
bam.mjlab.BamActuatorCfg.
Installation#
BAM is available on PyPI. Install it with the mjlab extra, which pulls in
mjlab together with mujoco, mujoco-warp and torch:
pip install better-actuator-models[mjlab]
Or, with uv:
uv add "better-actuator-models[mjlab]"
BAM is compatible with mjlab 1.3.
Registering the BAM init event#
First, be sure to register the bam_init function as a startup event in your environment:
from bam.mjlab import bam_init
# cfg is your mjlab task configuration
cfg.events["bam_init"] = EventTermCfg(func=bam_init, mode="startup")
Instantiating the config#
BamActuatorCfg is a dataclass that plugs into mjlab’s
actuator system. When an Entity is built, it instantiates a
BamActuator that runs the full BAM pipeline — voltage
control law, DC motor torque, and BAM friction budget — fully vectorized
over all parallel environments via PyTorch tensors.
Two approaches are available, mutually exclusive:
Bundled motor:
from bam.mjlab import BamActuatorCfg
actuator_cfg = BamActuatorCfg(
motor_name="{actuator}",
model="m6",
target_names_expr=(r".*",),
)
Custom JSON (parameters produced by bam.fit):
actuator_cfg = BamActuatorCfg(
json_path="params/my_motor/m6.json",
target_names_expr=(r".*",),
)
The target_names_expr field is a tuple of regex patterns that select
which actuated joints this config controls.
Supported bundled motors: see the list of identified actuators.
Supported model variants:
"m1"through"m6"(see Friction Models (M1-M6)).
Voltage and P-gain overrides#
By default, the supply voltage and firmware P-gain are read from the parameter JSON. They can be overridden at config level:
actuator_cfg = BamActuatorCfg(
motor_name="{actuator}",
model="m6",
target_names_expr=(r".*",),
vin=7.5, # supply voltage [V]
kp_fw=125.0, # firmware P-gain
)
Domain randomization#
BamActuatorCfg supports per-environment randomization
of two physical quantities that are naturally variable across hardware units
or charge states.
Battery voltage — sample a different supply voltage for each environment at startup:
actuator_cfg = BamActuatorCfg(
motor_name="{actuator}",
model="m6",
target_names_expr=(r".*",),
vin_range=(7.0, 8.0), # sampled uniformly at startup [V]
)
vin_range takes precedence over vin when both are set.
Voltage drop resistance — model battery + cable resistance with a per-env equivalent resistor between the battery and the motors:
where \(R_\text{drop}\) (vin_drop_resistance_range) is the combined
battery + wire resistance in ohms, \(\tau_i\) is the actuator torque on
joint \(i\), \(d_i\) its PWM duty cycle, and \(K_t\) the torque
constant. Randomizing this resistance captures variability in cable length or
connector quality across units:
actuator_cfg = BamActuatorCfg(
motor_name="{actuator}",
model="m6",
target_names_expr=(r".*",),
vin_range=(7.0, 8.0),
vin_drop_resistance_range=(0.05, 0.15), # [Ohm] ~100 mOhms of wire & battery resistance
vin_min=6.0, # hard lower bound [V]
)
Both ranges are sampled once at initialization and held constant across episode resets.
\(I\) is the current drawn from the battery, not the motor current: an
H-bridge in PWM behaves like a buck stage, so the motor draws
\(\tau_i / K_t\) continuously while the battery only sources it during the
PWM on-time, giving \(I_\text{bat} = d \, I_\text{motor}\). The product is
signed — a joint whose duty cycle and torque disagree in sign is braking and
returning current to the bus — and because all joints of a
BamActuator share one supply, the per-joint currents are
summed before the \(\max(0, \cdot)\). That clamp deliberately discards
regeneration raising the bus voltage, so the model stays conservative during
aggressive decelerations. Both the duty cycle and the torque are taken from the
previous solve, so the drop lags the load by one timestep.
This matches MujocoController exactly, so a policy trained
here sees the same battery behaviour as the CPU rollout — see
MuJoCo (CPU) for the same derivation.
Warning
The voltage drop is computed independently by each
BamActuator from its own joints’ current draw. If several
actuator configs share the same physical battery, their currents are not
summed together, so the modeled drop underestimates the real one. Group all
joints powered by the same battery under a single BamActuatorCfg
if you need the shared-supply behavior. In the current implementation, only one model
can be used per actuator config, voltage drop will not function properly if different motors
are mixed in the same config.
Command delay#
BAM inherits mjlab’s command delay system, which models the latency between policy output and motor response (e.g. communication bus latency, firmware scheduling). The lag is expressed in simulation steps and can be randomized per environment:
actuator_cfg = BamActuatorCfg(
motor_name="{actuator}",
model="m6",
target_names_expr=(r".*",),
delay_min_lag=1, # always at least 1 step of delay
# = 5ms with mjlab's default 200Hz control loop
delay_max_lag=3, # up to 3 steps, randomized per env
# = 15ms with mjlab's default 200Hz control loop
)
Setting delay_min_lag == delay_max_lag gives a fixed, deterministic delay.
Leave both at 0 (default) to disable delay entirely.
Passing the config to an Entity#
Pass the config to the actuator_cfgs argument of an mjlab Entity:
import mjlab
actuator_cfg = BamActuatorCfg(
motor_name="{actuator}",
model="m6",
target_names_expr=(r".*",),
kp_fw=125,
vin_range=(7.0, 8.0),
vin_drop_resistance_range=(0.05, 0.15), # [Ohm]
vin_min=6.0,
delay_min_lag=1,
delay_max_lag=3,
)
entity = mjlab.Entity(
xml_path="robot.xml",
actuator_cfgs=(actuator_cfg,),
)
mjlab calls build() internally to create the
BamActuator and wire it into the simulation graph.