[Question] Best practices for using the sysid module on floating-base bipedal humanoids?

[xialijun-sss]

Intro

Hi!
I am a researcher focusing on reinforcement learning and Sim-to-Real transfer for legged robots. I use MuJoCo extensively for robotic simulation, high-fidelity actuator modeling, and training locomotion policies.

My setup

• MuJoCo version: 3.5.x
• API: Python
• OS: Linux

My question

First of all, thank you for the 3.5 release! The new sysid module is exactly what the robotics community needs.

I am trying to use the sysid module to identify crucial dynamic parameters for a bipedal humanoid robot using hardware data. Specifically, I want to identify the base mass, Center of Mass (CoM), inertia tensors, alongside joint-level parameters.

The issue I encountered is that the current documentation and examples heavily focus on fixed-base robotic arms. Since a biped is a floating-base, underactuated system that relies entirely on intermittent contact dynamics, I am unsure how to correctly set up the optimization problem using the new tools.

Could you provide some guidance on the following:

1. Floating-Base & Contacts: Does the current sysid pipeline naturally handle floating-base kinematics (e.g., freejoint) and foot contacts during the residual computation, or do I need to explicitly handle contact states/forces in the loss function?
2. Future Examples: Are there any plans to release a legged robot sysid example in the near future?

[giulioturrisi]

Hi! You can see here https://github.com/iit-DLSLab/sim2real-robot-identification how we handle quadrupeds robot. We do only estimate leg parameters, without contact, with trajectories taken with the robot "in air"