Online Learning-Based Inertial Parameter Identification of Unknown Object for Model-Based Control of Wheeled Humanoids

TL;DR

This work proposes a fast, online learning-based inertial parameter estimation framework that enhances model-based control and addresses both parametric and non-parametric modeling errors independently using Robot System Identification and Gaussian Processes.

Abstract

Identifying the dynamic properties of manipulated objects is essential for safe and accurate robot control. Most methods rely on low noise force torque sensors, long exciting signals, and solving nonlinear optimization problems, making the estimation process slow. In this work, we propose a fast, online learning based inertial parameter estimation framework that enhances model based control. We aim to quickly and accurately estimate the parameters of an unknown object using only the robot's proprioception through end to end learning, which is applicable for real-time system. To effectively capture features in robot proprioception affected by object dynamics and address the challenge of obtaining ground truth inertial parameters in the real world, we developed a high fidelity simulation that uses more accurate robot dynamics through real-to-sim adaptation. Since our adaptation focuses solely on the robot, task-relevant data (e.g., holding an object) is not required from the real world, simplifying the data collection process. Moreover, we address both parametric and non-parametric modeling errors independently using Robot System Identification and Gaussian Processes. We validate our estimator to assess how quickly and accurately it can estimate physically feasible parameters of an manipulated object given a specific trajectory obtained from a wheeled humanoid robot. Our estimator achieves faster estimation speeds (around 0.1 seconds) while maintaining accuracy comparable to other methods. Additionally, our estimator further highlight its benefits in improving the performance of model based control by compensating object's dynamics and reinitializing new equilibrium point of wheeled humanoid

Figures (8)

• Figure 1: Conceptual overview of the proposed method. Thousands of wheeled humanoid robots, SATYRR, in a simulator shake an object to identify its inertial parameters. Different object dynamics impact the robot’s proprioception in varied ways. The effect of different object dynamics on the robot's proprioception closely mirrors real-world conditions due to effective real-to-sim adaptation. During training, the object's inertial parameters can be sampled either randomly or within specific shape boundaries.
• Figure 2: Customized Object To Get Ground-Truth Inertial Parameters. Depending on the location of the weights, the object can be represented as a barbell(3), hammer(8), etc. Right side image shows a sample of an object(3) with ground truth inertial parameters. The CoM is defined based on a central fixed coordinate system.
• Figure 3: Input Data Distribution in Training Dataset. The graphs shows the mean and standard deviation of joint trajectories in training dataset. This represents how different object's dynamics properties affect the trajectories of joint position and velocity for each joint.
• Figure 4: Results of Real2Sim Adaptation. The graphs depict the mean and standard deviation of the normalized Mean Squared Error (MSE) between trajectories derived from simulation and the real world, employing min-max normalization for straightforward comparison. The 45 target objects, which are not considered in the optimization process, are utilized for evaluation. The numbers on the graph represent the MSE. Based on the normalized total MSE outcomes, Sim+SysID+GP exhibits the smallest reality gap. The MSE for Sim+SysID is also notably low, showing only a minor performance gap compared to Sim+SysID+GP. This is because the error of parametric model has a large portion in causing the reality gap. The error from non-parametric modeling can become substantially larger in scenarios involving contact.
• Figure 5: Trajectory Comparison Results Obtained From Real World and Simulation. The trajectory produced by Puresim is closely located with the desired trajectory (black). First, SysID is applied to PureSim, followed by GP to SysID to handle the errors in parametric and non-parametric modeling, respectively. Each step brings the simulation trajectory closer to the target trajectory (red). In the case of Sim+SysID+GP, all joint trajectories in position and velocity are almost perfectly matched with the actual trajectories.