Reproduce “InEKFormer: A Hybrid State Estimator for Humanoid Robots” on OpenResearch

About this paper

Lasse Hohmeyer, Mihaela Popescu, Ivan Bergonzani, Dennis Mronga, Frank Kirchner

The InEKFormer, a hybrid state estimator developed by researchers at the University of Bremen and DFKI, integrates an Invariant Extended Kalman Filter (InEKF) with a Transformer network to enhance floating base state estimation for the RH5 humanoid robot. It demonstrates superior estimation accuracy over KalmanNet in short-sequence tests and benefits from increased training data, while also providing a new open-source dataset for humanoid robotics research.

Key insights

• Transformer networks are effective for learning complex statistical information required for Kalman gain prediction in high-dimensional humanoid robot state estimation, outperforming RNN-based approaches like KalmanNet.
• Autoregressive training, though challenging to stabilize for high-dimensional regression tasks, consistently yields superior performance compared to teacher-forcing mode.
• A significant sim-to-real gap exists, as models trained solely on simulated data perform substantially worse on real robot data, highlighting the necessity of real-world data or improved simulation fidelity for diverse motions.

Summary by alphaXiv.