A Latching Recovery Funnel for Neural Swing-Up Controllers of N -Link Cart–Pendulums
Abstract Neural swing-up controllers for multi-link cart–pendulum systems are typically optimized from a canonical initial condition in which the pendulum chain hangs at rest. We demonstrate that, although such controllers can successfully swing up and stabilize systems comprising (N=1)–(4) links from rest, their ability to recover from large disturbances is severely limited. Across 200 randomly generated knock-down states for each link configuration, the controller achieves a recovery rate of 97% for (N=1), but no more than 1% for (N\geq2). This degradation arises because the memoryless swing-up policy continues to inject control forces when the system enters states that lie substantially outside the trajectories encountered during training. To address this limitation, we introduce a recovery-funnel architecture integrated into the original feed-forward neural network through a compact, analytically designed control circuit. The proposed architecture comprises three functional components: (i) a downward-position linear quadratic regulator (LQR) damper that drives the cart towards its central equilibrium while bringing the pendulum links to rest; (ii) a detector neuron that identifies the downward-rest configuration; and (iii) a single self-recurrent rectified linear unit (ReLU) latch neuron that maintains activation of the swing-up control branch for a prescribed duration. The resulting architecture employs only ReLU and sinusoidal activation units, respects the original parameter bounds by keeping all weights below the prescribed limit, and preserves the nominal control behaviour. Simulation results demonstrate a substantial improvement in disturbance recovery, achieving successful recovery in all 200/200 trials for every configuration with (N\in{1,2,3,4}). Furthermore, the controller successfully recovers from all 24 extreme disturbance scenarios involving strong external pushes and impulsive tip-yank perturbations. These improvements come at the expense of increased recovery time, with median values ranging from 3.3 s for the single-link system to 12.6 s for the four-link configuration. The proposed recovery funnel relies on exact convergence to the downward-rest state in a noise-free simulation environment, which introduces important limitations regarding robustness to measurement noise, modelling uncertainties, and real-world implementation. These limitations are discussed alongside the implications of the proposed architecture. The source code, trained models, and evaluation framework are made openly available to facilitate reproducibility and further investigation.
Authors
- Mayank Goswami (ORCID: https://orcid.org/0000-0001-8829-9980)
- Shashank Kumar (ORCID: https://orcid.org/0000-0002-6456-8565)
Institutions
- Indian Institute of Technology Roorkee (IN)
Publication Details
- Journal
- Research Square
- Published
- 2026-11-17
- DOI
- https://doi.org/10.21203/rs.3.rs-11739298/v1
- Primary Topic
- Neural Networks and Applications
- Type
- preprint