Online Learning of Heterogeneous Locomotion Dynamics under Reinforcement Learning Policies for Precise Loco-Manipulation
Online Learning of Heterogeneous Locomotion Dynamics under Reinforcement Learning Policies for Precise Loco-Manipulation
This research addresses the challenge of reusing pretrained locomotion policies for precise loco-manipulation by introducing an online method to learn closed-loop dynamics under frozen policies. The authors propose a contaminated Gaussian model combined with Gaussian processes and random Fourier features to approximate heterogeneous responses caused by unobserved latent conditions. By weighting online updates based on the probability of predictable responses, the method reduces bias and fluctuations in dynamics estimates. Model predictive control (MPC) is then used for accurate floating-base motion prediction. The study establishes a finite-sample error bound ensuring uniform convergence and demonstrates consistent end-effector tracking with errors of 0.01 m across various manipulator configurations and tasks through both simulation and hardware experiments.
The paper proposes an online learning framework using a contaminated Gaussian model and Gaussian processes to handle heterogeneous locomotion dynamics under frozen policies, achieving end-effector tracking errors of 0.01 m in simulations and hardware experiments across different manipulator configurations.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
end-effector tracking error
0.01
Gaussian components
2
View original evidence
meaning that the same observed state will lead to different motions. Direct learning without accounting for this heterogeneity introduces bias and fluctuations into dynamics estimates, degrading floating-base prediction and end-effector tracking. To address this problem, a contaminated Gaussian model is proposed to approximate the effects of latent conditions. The model uses two Gaussian components to approximate pre
Open source S3
- paper: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
h errors of 0.01 m.
Open source S6
heterogeneous observations. With consistent and unbiased dynamics estimates, model predictive control (MPC) predicts floating-base motion accurately and enables precise loco-manipulation. A finite-sample error bound is established to ensure uniform convergence in online learning. Simulation and hardware experiments demonstrate consistent end-effector tracking across different manipulator configurations and tasks, wit
Open source S5
arXiv:2610.41892v1 Announce Type: new Abstract: How to reuse a pretrained locomotion policy for precise loco-manipulation is still an open problem. We propose an online method to learn the closed-loop dynamics under a frozen locomotion policy. The learned model predicts floating-base motion and improves loco-manipulation accuracy. However, unobserved latent conditions of the policy introduce heterogeneous responses,
Open source S2
meaning that the same observed state will lead to different motions. Direct learning without accounting for this heterogeneity introduces bias and fluctuations into dynamics estimates, degrading floating-base prediction and end-effector tracking. To address this problem, a contaminated Gaussian model is proposed to approximate the effects of latent conditions. The model uses two Gaussian components to approximate pre
Open source S3
Source：arXiv Robotics — research abstracts · arxiv.org