An Effectiveness Study Across Baseline and Learning-based Force Estimation Methods on the da Vinci Research Kit Si System

TL;DR

The paper tackles the absence of direct force sensing in early da Vinci–based surgical platforms by extending a learning-based force estimation method to the newer dVRK-Si and benchmarking it against baseline, torque-approximation methods. Using an LSTM to predict free-space joint torques from kinematics, the approach computes tip forces via $\hat{F}_{ext} = J^{-T}(\tau - \hat{\tau}_{LSTM})$, and is evaluated on ~50 minutes of teleoperation data with a separate tip-contact test. Results show the learning-based estimator achieves an average RMSE of $5.21\%$ on dVRK-Si, outperforming all baselines and comparable to prior dVRK Classic performance, though dVRK-Si remains 2–3× less accurate than the Classic due to poorer PID control and gravity compensation. The study identifies gravity compensation as a key factor driving residual errors on dVRK-Si and presents the first system identification of this platform, informing design and sensing strategies for robust sensorless force estimation in surgical robotics.

Abstract

Robot-assisted minimally invasive surgery, such as through the da Vinci systems, improves precision and patient outcomes. However, da Vinci systems prior to da Vinci 5, lacked direct force-sensing capabilities, forcing surgeons to operate without the haptic feedback they get through laparoscopy. Our prior work restored force sensing through machine learning-based force estimation for the da Vinci Research Kit (dVRK) Classic. This study extends our previous method to the newer dVRK system, the dVRK-Si. Additionally, we benchmark the performance of the learning-based algorithm against baseline methods (which make simplifying assumptions on the torque) to study how the two systems differ. Results show the learning-based method achieves an average root-mean-square-error (RMSE) of 5.21\%, for the dVRK-Si, which is comparable to the dVRK Classic. In both systems, the learning-based method outperforms baselines, but the difference is much larger in the dVRK-Si. Nonetheless, dVRK-Si force estimation accuracy lags behind the dVRK Classic, with RMSE 2 to 3 times higher. Further analysis reveals poor PID control in the dVRK-Si. We hypothesize that this is due to the lack of gravity compensation, as unlike the dVRK Classic, the dVRK-Si is not mechanically balanced. This study advances the understanding of learning-based force estimation and is the first work to characterize the dynamics of the new dVRK-Si system.

Figures (7)

• Figure 1: Experimental setup. From left to right are the dVRK Classic PSM arm, the F/T sensor with a 3D-printed shaft mounted on top, and the dVRK-Si PSM arm.
• Figure 2: LSTM network architecture for one joint. Six identical networks are trained for each joint to account for different ranges of motion Wu2021RobotFE.
• Figure 3: Comparison between pre- and post-filtered force estimation results on the x-axis. The blue curves represent the ground truth from the ATI F/T sensor. The yellow and red curves represent the estimated results.
• Figure 4: The baseline and learning-based method force estimation results on dVRK Classic (top) and dVRK-Si (bottom), alongside ground truth force measurements. The baseline estimations perform much worse on dVRK-Si than on dVRK Classic, as they all deviate considerably from the ground truth curve. For the X-axis, baseline estimation errors on dVRK-Si can reach 30-40 N.
• Figure 5: Comparison of transient joint torque response between the dVRK Classic and dVRK-Si. Both arms are set to home position and given a 0.1 Nm desired torque command at 30 ms. Frequency was set to 100 Hz. The dVRK Classic takes 30 ms to reach steady state, while the dVRK-Si achieves steady state within 10 ms, matching the command timing without delay. This shows the torque response of dVRK-Si outperforms dVRK Classic. Note that dVRK Classic is self-balanced and requires no torque at the initial position, while dVRK-Si needs about -1.47 Nm to compensate for gravity.