Your robot sim is lying about friction

What it does

BAM is a Python toolkit that fits extended friction models to real servo trajectories, then drops them into MuJoCo. It ships ready-made models for seven common actuators—including Dynamixel and eRob series—so you can swap out the default Coulomb-Viscous approximation for something that acknowledges Stribeck effects and load-dependent drag.

The interesting bit

Most simulators treat servo friction like a simple dimmer switch; BAM treats it like the weird mechanical mess it actually is. The payoff is a narrower sim-to-real gap when you are doing reinforcement learning on physical hardware.

Key highlights

• Pre-identified models for seven servos (Dynamixel MX-64/106/XL-320/XL330, eRob80:50/100, Feetech STS3215)
• Supports both MuJoCo CPU and the newer MuJoCo Warp backend
• Includes an identification pipeline to fit models from your own recorded trajectories
• Targets complex phenomena like Stribeck, load-dependence, and quadratic friction
• Backed by an ICRA 2025 paper

Caveats

• The README is essentially a landing page; most technical details live on ReadTheDocs.
• Only seven servos have pre-built models; anything else requires running the identification pipeline yourself.
• The BibTeX software citation lists the release as version 0.0.1.

Verdict

Grab this if you are training RL policies on real robots and wondering why your sim-perfect gait falls apart on hardware. Skip it if your actuators already behave like idealized physics homework.