Physics-Informed LSTM-Based Delay Compensation Framework for Teleoperated UGVs

TL;DR

This paper tackles latency in bilateral teleoperation of low-speed UGVs navigating soft terrains by introducing a physics-informed LSTM (PiLSTM) predictor framework that combines data-driven learning with physical constraints. Four predictors (two forward, two backward) are integrated into a blended bilateral control to compensate large network delays, restoring high-fidelity command-tracking and improving transparency. Open-loop results show an average delay-compensation improvement of about 26.1% over conventional predictors, while human-in-the-loop experiments confirm enhanced closed-loop performance and reduced task times in lunar-like soft-terrain scenarios. The work advances teleoperation for planetary exploration by delivering robust delay compensation under nonlinear, time-varying wheel-terrain interactions, with significant implications for safety and efficiency in remote UGV operation.

Abstract

Bilateral teleoperation of low-speed Unmanned Ground Vehicles (UGVs) on soft terrains is crucial for applications like lunar exploration, offering effective control of terrain-induced longitudinal slippage. However, latency arising from transmission delays over a network presents a challenge in maintaining high-fidelity closed-loop integration, potentially hindering UGV controls and leading to poor command-tracking performance. To address this challenge, this paper proposes a novel predictor framework that employs a Physics-informed Long Short-Term Memory (PiLSTM) network for designing bilateral teleoperator controls that effectively compensate for large delays. Contrasting with conventional model-free predictor frameworks, which are limited by their linear nature in capturing nonlinear and temporal dynamic behaviors, our approach integrates the LSTM structure with physical constraints for enhanced performance and better generalization across varied scenarios. Specifically, four distinct predictors were employed in the framework: two compensate for forward delays, while the other two compensate for backward delays. Due to their effectiveness in learning from temporal data, the proposed PiLSTM framework demonstrates a 26.1\ improvement in delay compensation over the conventional model-free predictors for large delays in open-loop case studies. Subsequently, experiments were conducted to validate the efficacy of the framework in close-loop scenarios, particularly to compensate for the real-network delays experienced by teleoperated UGVs coupled with longitudinal slippage. The results confirm the proposed framework is effective in restoring the fidelity of the closed-loop integration. This improvement is showcased through improved performance and transparency, which leads to excellent command-tracking performance.

Figures (16)

• Figure 1: The conceptual relationship illustrates how the command-tracking performance of teleoperated UGV is affected by both the terrain properties and the latency.
• Figure 2: Proposed predictor framework, comprising two model-free predictors namely; forward predictor and backward predictor, and a blended architecture for the teleoperated UGV subjected to terrain-induced slippage. The Red lines represent the Delayed case, and the Green Lines represent the Predicted case.
• Figure 3: Physics-informed LSTM architecture for the proposed predictor: Number of fixed-length input features $n$, number of dense layer hidden units $m$, number of LSTM layers $k$, number of LSTM layer hidden units $l$, and a single-step target output at $n$.
• Figure 4: Human-in-the-loop test platform for teleoperated UGV traversing soft terrain with slippage awareness.
• Figure 5: Designated soft terrain tracks A, B, and C; UGV, and non-track zone in the virtual environment.