BAM: Better Actuator Models

Project notes

Use the dependency set for your task

The reviewed pyproject.toml declares package version 1.0.3 and Python >=3.12,<3.13, with numpy and colorama as base dependencies. This differs from the README's historical software-citation version 0.0.1; the citation is not the current package declaration. Select the source revision and compatible environment rather than combining examples from different releases. No installation was reproduced. [1] [3]

For the CPU route, the guide documents pip install 'better-actuator-models[mujoco]'. GPU/mjlab integration and physical identification use separate extras. The reviewed mjlab extra constrains mjlab to 1.3, MuJoCo and MuJoCo Warp to the 3.7 series and warp-lang to the 1.12 series; source comments explain API incompatibilities with newer stacks. Installing every extra is unnecessary for an initial CPU model investigation. [3] [6]

Match the motor and MJCF representation

The CPU guide begins with bam.model.load_model, selecting a bundled motor name and model variant m1 through m6, or a custom parameter JSON. The placeholder {actuator} needs a supported motor selection; it is not a literal working model name. Inspect the identified-actuator list referenced by the guide and match the physical servo and identification conditions before interpreting predictions. [5] [6]

Its MJCF requirement is concrete: actuators must be declared as motor, rather than position or velocity actuators. BAM overwrites frictionloss, damping and armature at runtime, so values placed in those XML fields are not preserved. The documented CPU controller computes motor torque, writes control and friction quantities, and can model load-dependent supply-voltage drop. These are upstream interface descriptions, not observed simulation results. [6]

The inspected actuator source also distinguishes stateful models: they must be evaluated once per timestep in chronological order, and independent simulations sharing a model require separate saved/restored state. That matters when moving a single-model experiment into parallel environments; the excerpt does not verify a particular batching integration. [4]

The repository and package declare Apache-2.0. Simulator dependencies, hardware drivers and any external recordings or models retain separate terms. The identification extra includes hardware/optimization/logging dependencies; installing it does not collect calibration data or validate a physical testbench. No dependency install, identification run, CPU/GPU simulation, friction fit, accuracy benchmark or sim-to-real transfer test was performed. [2] [3]