Multi-task Gaussian Process Learning of Robot Inverse Dynamics

Kian Ming Adam Chai, Christopher K. I. Williams, Stefan Klanke, Sethu VijayakumarNeurIPS 2008 · OpenReview

Introduction

In robotics, inverse dynamics is the critical process of calculating the joint torques required to achieve a specific motion, defined by position, velocity, and acceleration. While analytical physics models exist for these calculations, they often fail in real-world applications due to complex friction, cable tension, or uncertain physical parameters. Consequently, researchers frequently turn to machine learning to empirically "learn" these dynamics.

However, a major challenge arises when a robot interacts with its environment, such as picking up different objects. Each new load fundamentally changes the robot's dynamics. Learning a completely new model for every possible object is highly data-inefficient, while simply "pooling" all data into one giant model ignores the specific physical nuances of different weights. This paper addresses this challenge by treating it as a multi-task learning (MTL) problem, where each "task" corresponds to the robot operating with a different load.

The Multi-Task Learning Approach

The authors propose a Multi-task Gaussian Process (mGP) model grounded in a key physical insight: for a rigid-body manipulator, the torque at a joint is a linear function of the inertial parameters (like mass and center of gravity). Crucially, when a robot picks up a load, only the inertial parameters of the last link are modified.

To exploit this, the researchers developed a model with several key technical components:

  • Linear Decomposition: The torque for a joint is decomposed into a set of shared latent functions (representing the robot's inherent structure) and task-specific weights (representing the load's unique inertial properties).
  • Covariance Structure: The model utilizes a shared covariance function across inputs (the trajectory) and a positive semi-definite "task similarity matrix" that captures how related different loads are to one another.
  • Staged Optimization: Training the model involves a three-step heuristic: optimizing kernel parameters for a single task, estimating an initial task similarity matrix, and performing a final joint optimization of all hyperparameters.
  • Model Selection: The Bayesian Information Criterion (BIC) is used to determine the optimal rank of the similarity matrix. This prevents overfitting while naturally capturing the underlying 11-dimensional space of inertial parameters.

Experimental Setup

The study validates this approach using data generated from a realistic simulation of a PUMA 560 robot arm, which features 6 degrees of freedom.

Diagram of the PUMA 560 robot arm, illustrating the 6 degrees of freedom.

The experimental dataset encompasses:

  • Loads: 15 different loads, ranging from 0.2kg to 3.0kg, modeled as cuboids with varying dimensions and attachment points.
  • Trajectories: Four distinct figure-of-eight paths executed at four different speeds.
  • Testing Scenarios: The model is evaluated on interpolation (predicting torques for seen trajectories with seen loads) and extrapolation (predicting torques for entirely unseen trajectories or novel load/path combinations).

Key Insights and Results

The most powerful insight from this research is that the structure of the multi-task GP is not just a statistical convenience—it is a direct reflection of the robot's physical equations. By sharing the "latent functions" across all loads, the robot learns the general "feel" of its own limbs from all available data, even if it has only seen a specific load on a very limited path.

Performance comparison on Joint 1 for interpolation and extrapolation tests.

The multi-task GP (specifically the mGP-BIC variant) consistently outperformed baselines, including single-task GPs, independent GPs, and pooled GPs.

While simple models perform adequately when interpolating known paths, the multi-task GP truly shines in extrapolation. In the hardest test cases—extrapolating to unseen trajectories—the mGP-BIC achieved significantly lower normalized Mean Square Errors (nMSE). For instance, on joints like Joint 1 and Joint 4, the mGP-BIC maintained high accuracy in regimes where other models' errors spiked dramatically. Furthermore, the model proved highly data-efficient, reaching high accuracy with far fewer samples per task than independent models require, thanks to effective information transfer between different loading contexts.

Limitations and Future Directions

Despite its strong performance, the approach has limitations:

  • Computational Complexity: Like most standard GP-based methods, this approach suffers from $O(N^3)$ scaling. While the authors use Kronecker-product structures to speed up calculations for complete datasets, scaling to massive real-time datasets will likely require sparse GP approximations.
  • Independent Joints: Currently, the model treats each joint as an independent multi-task problem. Because robot joints are physically coupled, future work could investigate learning these couplings across joints to further improve accuracy.
  • Simulation vs. Reality: The simulation included realistic elements like friction, but real-world hardware introduces non-stationary noise and non-rigid body effects (such as cable stretch) that could challenge the linear parameter assumption underlying the model.

Summary updated

Abstract

The inverse dynamics problem for a robotic manipulator is to compute the torques needed at the joints to drive it along a given trajectory; it is beneficial to be able to learn this function for adaptive control. A given robot manipulator will often need to be controlled while holding different loads in its end effector, giving rise to a multi-task learning problem. We show how the structure of the inverse dynamics problem gives rise to a multi-task Gaussian process prior over functions, where the inter-task similarity depends on the underlying dynamic parameters. Experiments demonstrate that this multi-task formulation generally improves performance over either learning only on single tasks or pooling the data over all tasks.
BibTeX
@inproceedings{chai2008multi,
  author = {Kian Ming Adam Chai and Christopher K. I. Williams and Stefan Klanke and Sethu Vijayakumar},
  title = {{Multi-task Gaussian Process Learning of Robot Inverse Dynamics}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2008},
  url = {https://openreview.net/forum?id=kliogIZ99RM},
}
Figure