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
What problem does this paper attempt to address?
The problem that this paper attempts to solve is how to quickly and accurately identify the dynamic parameters (such as mass, center - of - mass position, moment of inertia, etc.) of unknown objects when a robot manipulates them, in order to improve the performance of model - based control. Specifically, existing methods usually rely on low - noise force - torque sensors, long - duration excitation signals, and solving nonlinear optimization problems, which makes the estimation process slow and complex. This paper proposes a dynamic parameter estimation framework based on online learning, aiming to use only the robot's proprioception (i.e., internal sensor data such as joint positions and velocities), and through an end - to - end learning method, achieve fast and accurate dynamic parameter estimation, which is suitable for real - time systems.
### Main Contributions
1. **Fast Dynamic Parameter Estimation Framework**: Use an end - to - end data - driven model, relying only on the robot's proprioceptive error, without the need for long excitation signals and constrained nonlinear optimization, to achieve fast dynamic parameter estimation.
2. **Effectively Handle Parameter and Non - Parameter Modeling Errors**: Independently handle parameter and non - parameter modeling errors through robot system identification (SysID) and Gaussian processes (GPs) to reduce the reality gap.
3. **Verification on Physical Hardware**: Demonstrate the benefits of accurate dynamic parameters for improving the performance of model - based controllers on physical hardware.
### Method Overview
1. **Real - to - Sim Adaptation**:
- **Offline Robot System Identification (SysID)**: Minimize the difference between simulation and real data by optimizing the robot system parameters.
- **Non - Parameter Dynamic Modeling (GPs)**: Use Gaussian processes to model non - parameter modeling errors, such as nonlinear friction and hysteresis.
2. **Learn the Dynamic Parameters of Unknown Objects**:
- **Dynamic Trajectory Planning and Manipulation Control**: Design periodic excitation trajectories and use the Hanning window function for smooth transitions, and control the robotic arm through inverse kinematics.
- **Database Construction**: Use the IsaacGym simulator to generate training data sets containing different objects.
- **Time - Series Data - Driven Model Training**: Use models such as 1D - CNN for training, and ensure the physical consistency of the estimated parameters through regularization terms.
### Experimental Verification
1. **Real - to - Sim Adaptation Verification**: Verify the effectiveness of SysID and GPs in reducing the reality gap by comparing different baseline methods.
2. **Dynamic Parameter Estimation Verification**: Compare with traditional methods (such as ordinary least squares and weighted least squares) and learning - based methods to evaluate the estimation speed and accuracy.
3. **Control Task Experiments**: Verify the effect of the estimated dynamic parameters on improving the model control performance in actual tasks, including object manipulation and transportation tasks.
### Results Discussion
- **Real - to - Sim Adaptation**: SysID and GPs significantly reduce the reality gap of position and velocity trajectories, especially when dealing with non - parameter modeling errors in contact scenarios.
- **Dynamic Parameter Estimation**: The proposed framework is superior to traditional methods in terms of estimation speed and accuracy, and performs well in actual tasks, significantly improving the model control performance.
In conclusion, this paper proposes an efficient dynamic parameter estimation method, which not only performs well in the simulation environment but also has high practical value in practical applications.