Record interaction
Collect the robot’s applied forces and torques together with the object’s observed pose trajectory.
RESEARCH PROJECT / IROS 2026
RigPIDynamic Parameter Identification of Rigid Body via VLM-Seeded Differentiable Simulation
Recovering physical parameters from real robot interactions through vision-language priors and differentiable physics.
RIGPI PIPELINE01 — 03
01
OBSERVEReal-world interactionMotion · Force · Torque
02
INITIALIZEVLM-seeded priorInformed start · Feasible bounds
03
REFINEDifferentiable simulationGradient-based optimization
IDENTIFIED PARAMETERSθ = { m, c, I, μ }
PHYSICAL PARAMETER IDENTIFICATIONSCROLL TO EXPLORE
01 THE IDEA
From interaction
to physical insight.
A robot can see an object move and feel the forces it applies. RigPI turns those observations into a physically grounded model.
Identifying mass, center of mass, inertia, and friction from real-world data is difficult when measurements are noisy and the starting guess is poor. RigPI combines vision-language model priors with a differentiable simulator to initialize, constrain, and refine these parameters.
mMassHow heavy
cCenter of massWhere weight acts
IInertiaHow it rotates
μFrictionHow it resists motion
02 FRAMEWORK
A closed loop between
reality and simulation.
RigPI uses recorded interaction data to refine a simulator until its predicted trajectory agrees with the observed motion.
03 HOW IT WORKS
Three stages.
One consistent model.
Semantic priors help the optimization start in a plausible region; physics and observed motion guide the final estimate.
Establish a prior
Use visual semantic cues to initialize physical properties and define a feasible parameter range.
Fit the simulation
Differentiate through the simulated trajectory and update parameters to reduce the observed pose error.
04 PROJECT VIDEO
See RigPI
in motion.
Watch the project video for the real-world setup and trajectory reproduction.
Open video in Drive05 PUBLICATION
Read more.
Build on it.
RigPI is an IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) 2026 paper.
BIBTEX
@article{he2026rigpi,
title={RigPI: Dynamic Parameter Identification of Rigid Body via VLM-Seeded Differentiable Simulation},
author={He, Xincheng and Zhang, Rongrong and Jiang, Wei and Xu, Wenqiang},
journal={arXiv preprint arXiv:2606.25212},
year={2026}
}