A few hundred milliseconds can matter in telesurgery. When a surgeon’s command and the robot’s response are separated by network delay, even small timing errors can reduce precision, disrupt feedback, and increase risk during delicate procedures. The article notes that delays as low as 200 ms, and even 150 ms in fine motor tasks, can impair surgical performance, making network latency a major barrier to reliable remote surgery. This problem is especially important for microsurgical or catheter-based teleoperation, where small deviations may affect the accuracy of tool movement. To address this problem, the authors developed a deep reinforcement learning framework that estimates both the direction and value of future network latency, aiming to give telesurgical systems earlier warning of changing communication conditions.
Unlike conventional regression-only approaches that focus mainly on predicting an exact delay value, the proposed method uses a two-stage forecasting strategy. First, a deep Q-network agent predicts the binary direction of latency change, classifying whether the next delay is likely to increase or decrease. Second, that directional signal is transformed into a quantitative delay estimate. This is important because, in a teleoperation setting, knowing whether latency is rising or falling can be as useful as knowing the predicted value itself. Directional information could help downstream control systems prepare for worsening delay, adjust compensation strategies, or adopt more conservative operating modes before the network condition becomes critical.
The study used latency data collected from simulated teleoperation sessions between an operator console and a robotic navigation platform. These sessions were designed to mimic the control traffic patterns of magnetically guided telesurgery systems, including three-dimensional orientation vectors, advance and retract commands, and timing metadata. Data were collected across an international path from Pasig City, Philippines, to Palapye, Botswana, as well as domestic Botswana network paths, allowing the dataset to capture variable delay, jitter, and packet-loss behavior. The authors then transformed raw latency measurements into engineered temporal indicators describing short-term fluctuations, medium-term trends, volatility, momentum, acceleration, and relative position within recent latency ranges. This feature-engineering process helped the model distinguish routine network noise from more sustained delay trends.
The model was trained and evaluated using a 70:15:15 split for training, validation, and testing. To preserve temporal causality, the feature and labeling strategy used only information available up to the current time, avoiding look-ahead bias. On the held-out test set, the deep Q-network achieved 85.8% overall direction prediction accuracy, a 35.8% improvement over a random classifier. It also outperformed baseline approaches such as a majority-class predictor and a simple moving-average crossover strategy. For quantitative latency estimation, the two-stage framework achieved a mean absolute percentage error of 7.23% and a value accuracy of 92.55%, showing that directional prediction can be converted into useful delay-value forecasting. These results indicate that the method can provide both interpretable trend information and numerical estimates for latency mitigation.
The results also reveal an important deployment challenge. When the model trained on the Philippines–Botswana network path was tested on different domestic Botswana paths without fine-tuning, its performance decreased because of distribution shifts between network environments. This suggests that real-world use will require domain adaptation, mixed-domain training, online fine-tuning, and safety-aware validation. Looking forward, the study points toward latency-aware telesurgical systems that can anticipate network changes before they disrupt surgical control. Future work may include hardware-in-the-loop testing with actual telesurgery robots, broader validation across diverse network paths, and safety mechanisms such as confidence monitoring, predictive display overlays, reduced teleoperation gain during unstable periods, and conservative fallback command policies. If validated in real robotic platforms, this direction-aware forecasting approach could help remote surgical systems become more resilient under unpredictable communication conditions.
SmartBot
Experimental study
Network Latency Estimation for Telesurgery Using Deep Reinforcement Learning
21-May-2026
The authors declare no conflicts of interest.