AI-MOLE: Autonomous Iterative Motion Learning for Unknown Nonlinear Dynamics with Extensive Experimental Validation

TL;DR

AI-MOLE addresses autonomous motion learning for systems with unknown nonlinear dynamics by iteratively applying input trajectories, building Gaussian Process-based models of the plant, and updating inputs with norm-optimal ILC until tracking improves. It supports both input/output only (IO) and state-aware (IS) settings, with autonomous parameterization enabling plug-and-play operation. The method is validated on three real-world robots across nine tracking tasks, achieving rapid learning within 5–10 trials per task and demonstrating faster convergence when state information is available. This work highlights how a GP-ILC hybrid can enable true autonomy in robotic learning without task- or system-specific prior knowledge.

Abstract

This work proposes Autonomous Iterative Motion Learning (AI-MOLE), a method that enables systems with unknown, nonlinear dynamics to autonomously learn to solve reference tracking tasks. The method iteratively applies an input trajectory to the unknown dynamics, trains a Gaussian process model based on the experimental data, and utilizes the model to update the input trajectory until desired tracking performance is achieved. Unlike existing approaches, the proposed method determines necessary parameters automatically, i.e., AI-MOLE works plug-and-play and without manual parameter tuning. Furthermore, AI-MOLE only requires input/output information, but can also exploit available state information to accelerate learning. While other approaches are typically only validated in simulation or on a single real-world testbed using manually tuned parameters, we present the unprecedented result of validating the proposed method on three different real-world robots and a total of nine different reference tracking tasks without requiring any a priori model information or manual parameter tuning. Over all systems and tasks, AI-MOLE rapidly learns to track the references without requiring any manual parameter tuning at all, even if only input/output information is available.

Figures (6)

• Figure 1: The method proposed in this work, called AI-MOLE, enables robotic systems with unknown, nonlinear dynamics to autonomously learn to track reference trajectories that lead to agile and dynamic motions. Learning only requires 5--15 trials, works in real-world environments and without any manual parameter tuning.
• Figure 2: AI-MOLE is an iterative learning scheme, in which each iteration consists of three steps. (A) The current-trial input trajectory is applied to the unknown dynamics yielding either an output or state trajectory. (B) The experimental data are then used to model the unknown dynamics using GPs. (C) Finally, the GP model is utilized by an ILC update rule to compute the next-trial input trajectory. The iterations continue until the output trajectory converges sufficiently close to the reference trajectory.
• Figure 3: The experimental evaluation of AI-MOLE consists of three real-world testbeds, namely, the balancing cube (CUBE), the two-wheeled inverted pendulum robot (TWIPR), and the double pendulum (PENDU). For each of the systems, AI-MOLE has to learn to track three different reference trajectories.
• Figure 4: The output trajectories for learning the respective first reference for CUBE, TWIPR, and PENDU show that for each task the initial output trajectory is close to zero and significantly deviates from the reference. From there onwards, the output trajectories quickly converge to the reference independent of the system, trial length, and the references' amplitudes and frequencies.
• Figure 5: Experimental evaluation of AI-MOLE (IO-version) on three testbeds and, for each testbed, three respective reference trajectories. Despite varying dynamics and references, and without any manual tuning, satisfying tracking performance is achieved within 5--10 trials in all nine tasks. The results demonstrate that AI-MOLE is capable of autonomously solving reference tracking tasks for a variety of different systems under real-world settings in a truly plug-and-play fashion.