Simulation-Augmented Hysteresis Compensation in Continuum Surgical Robots Via Residual Learning
Haolin Jiang, Thiusius R. Savarimuthu, Di Wu
AI summary
Problem
Continuum surgical robots suffer from tip positioning errors due to unmodeled nonlinear hysteresis, while purely data-driven compensation methods require extensive real-world training data that risks system wear and is time-consuming.
Approach
The method first computes residual actuation corrections to align a finite-element digital twin with sparse real data, then generates large-scale synthetic trajectories to train a deep learning inverse kinematics controller on a hybrid dataset.
Key results
- Achieves tip positioning errors of 1.344–3.678 mm on unseen trajectories
- Reduces normalized trajectory error to 2.0–5.3% under large bending
- Outperforms nominal model-based controllers in precision and hysteresis compensation
- Enables accurate open-loop control using only sparse real robot data
Why it matters
Provides a scalable, wear-free pathway for precise continuum robot control in minimally invasive surgery by bridging simulation and reality with minimal physical data.
Abstract
Continuum robots show promise for minimally invasive surgery, but suffer from tip positioning errors due to hysteresis. Pure analytical models struggle to capture nonlinear effects, while data-driven methods typically demand extensive amounts of robot training data, potentially leading to increased system wear. To overcome the limitations of both analytical and data-driven methods, we propose a novel Real2Sim2Real framework that establishes a Finite-Element (FE) digital twin of a continuum robot. The framework is structured in two stages. We first collect a small amount of real hysteresis data and identify the tendon-space residual actuation displace- ment required to minimize the pose error between the finite- element (FE) model and the physical robot. These residual displacements are then applied as auxiliary tendon inputs to enhance the realism of the simulation, ensuring that the hysteresis behavior of the digital twin aligns with that of the physical system. This enables the generation of a large-scale, high-fidelity synthetic dataset at minimal cost, substantially expanding the available training data. In the second stage, we train a deep-learning (DL)–based inverse kinematics model that maps desired tip poses to primary tendon displacements, using a hybrid dataset composed of both real and simulated data. The key advantage of this approach is that it achieves a high-accuracy controller while requiring only a small amount of real robot data. The trained controller was evaluated on various 2D trajectories. Results demonstrate effective hysteresis compensation, achieving tip positioning errors between 1.344 mm and 3.678 mm on unseen trajectories, even under large bending conditions. When normalized by the trajectory range, the error falls between 2.0% and 5.3%. Compared to a nominal model-based controller, the proposed simulation-augmented inverse controller achieves substantially higher precision and superior hysteresis compensation in continuum surgical robots.