Rhoban/bam

What it solves

BAM addresses the gap between simulation and reality in robotics by providing more accurate models of servo actuators. Standard simulators like MuJoCo and IsaacGym often use simplistic friction models (Coulomb-Viscous) that fail to capture complex phenomena such as Stribeck effects, load-dependence, and quadratic effects, which can hinder the transfer of Reinforcement Learning (RL) policies from simulation to real robots.

How it works

BAM improves actuator simulation through three main components:

1. Identification Pipeline: A system to fit friction models based on recorded trajectories.
2. Extended Friction Models: A set of models that capture complex friction behaviors beyond basic simulation defaults.
3. Pre-identified Models: A library of ready-to-use models for common servos, including various Dynamixel and Feetech models.

It provides a simple API to integrate these advanced models into MuJoCo CPU and MuJoCo Warp.

Who it’s for

Robotics researchers and engineers who use Reinforcement Learning to train policies on simulated robots and need high-fidelity actuator models to ensure successful sim-to-real transfer.

Highlights

• Accurate simulation of complex friction (Stribeck, load-dependence, quadratic effects).
• Identification pipeline for fitting models from real-world data.
• Pre-configured models for common servos like Dynamixel and eRob80.
• Integration with MuJoCo CPU and MuJoCo Warp.