Abstract
Learning the inverse dynamics of robotic systems is essential for precise control but challenging due to nonlinearities, non-stationary data streams, and the need for real-time uncertainty estimation. We present Infinite Local Regression (ILR), a Bayesian nonparametric approach that employs Dirichlet Process Mixtures of local linear experts. By utilizing a Variational Bayes inference scheme, ILR achieves high-frequency online control while automatically scaling model complexity and providing robust uncertainty quantification.
Introduction
In robotics, inverse dynamics refers to the mapping from desired joint positions, velocities, and accelerations to the required motor torques. Accurate modeling is difficult because of unmodeled physical effects like friction and cable tension. Furthermore, robots must learn incrementally from data streams without experiencing "catastrophic forgetting." Existing solutions like Gaussian Process Regression (GPR) provide excellent accuracy and uncertainty estimates but suffer from poor scalability. Local Regression (LR) methods offer the necessary speed but often rely on heuristic-based model management. ILR addresses these limitations by providing a principled, scalable, and probabilistic framework for dynamics learning.
Methodology: Infinite Local Regression
The core of ILR is the treatment of robot dynamics as a mixture of many local linear "experts."
As illustrated in Figure 4, the model employs a Dirichlet Process with a stick-breaking construction, allowing it to support an infinite number of components theoretically. New experts are allocated only as the complexity of the data demands. Unlike standard regression, ILR models the joint distribution of inputs and outputs, which naturally defines "receptive fields"—regions of the state space where specific experts are valid. To maintain the high-frequency requirements of real-time control, the authors utilize Variational Inference (specifically a VBEM algorithm) to derive closed-form parameter updates, avoiding the computational overhead of sampling-based methods.
Experimental Results and Discussion
The ILR model was validated on synthetic functions and real-world datasets, including the 7-DoF SARCOS anthropomorphic arm and a 4-DoF cable-driven Barrett-WAM manipulator.
Principled Complexity and Uncertainty
Figure 3 demonstrates the model's ability to handle diverse data patterns. The Dirichlet Process prior acts as a regularizer, preventing over-fitting while ensuring sufficient model capacity. Crucially, the model provides reliable uncertainty estimates; when the robot enters an unseen state, the predictive uncertainty increases, and the mean prediction reverts to a safe prior. This "knowing what you don't know" is vital for safe autonomous exploration.
Real-Time Performance and Accuracy
On the Barrett-WAM, ILR achieved a prediction frequency of 2000 Hz, well above the 500 Hz threshold required for stable control. In trajectory-tracking tasks, the ILR-augmented controller reduced Mean Squared Error (MSE) by an order of magnitude compared to standard PD controllers. For instance, in one specific task, MSE dropped from $2.33 \times 10^{-2}$ (PD) to $1.03 \times 10^{-3}$ (ILR+PD).
Conclusion
Infinite Local Regression offers a robust solution for learning robot dynamics by combining the flexibility of Bayesian nonparametrics with the efficiency of Variational Inference. It eliminates the need for manual tuning of model complexity and provides the real-time performance necessary for practical robotics applications. Future research will investigate hierarchical models to further enhance data efficiency through parameter sharing across local experts.
Summary updated