The Missing Sense in Robotic Surgery: A Low-Cost Fix That Doubles Training Success

Surgeons operating the most advanced robotic systems in the world are, in an important sense, flying blind. They can see in high-definition 3D. They can rotate their instruments 540 degrees. They can eliminate hand tremor entirely. But they cannot feel a thing. When a robotic surgeon grips tissue too hard, or presses a needle through a layer that should have given way, there is no tactile whisper to warn them. Only the image on the screen — and by the time that image shows the consequences, the damage is already done.

This missing sense is not a minor inconvenience. It is a root cause of suture breakage, tissue trauma, and surgical errors in robotic-assisted procedures. And it has become increasingly urgent as robotic surgery spreads to hospitals that are training the next generation of surgeons on these systems. A new paper from Heriot-Watt University in Edinburgh proposes a practical, low-cost solution: a modular surgical instrument that plugs into an existing training robot and restores haptic feedback — the sense of touch and force — to the trainee's hand. In a controlled user study, the result was striking. Success rates on a delicate force regulation task doubled, from 27% to 54%, when haptic feedback was switched on (Shaker & Erden, 2026).

The Science

The problem begins with the architecture of robotic surgery itself. In open surgery, a surgeon's hands are in direct contact with the patient. In laparoscopic surgery — the kind done through small keyhole incisions — long instruments attenuate but do not eliminate tactile sensation. In robotic surgery, the surgeon sits at a console, often across the room, manipulating joystick-like controls that command a robotic arm holding the instrument. That chain of mechanical translation strips out haptic information almost entirely.

The gold-standard commercial systems have started to address this. Intuitive Surgical's da Vinci 5, released in 2024, now includes force feedback. Asensus Surgical's Senhance system offers tactile feedback. But these platforms cost millions of dollars to acquire and maintain, placing them well beyond the reach of most training programs — especially in resource-limited settings. The result is a paradox: the very institutions that most need to train surgeons on the haptic nuances of robotic technique are the ones least likely to have equipment that teaches it.

Shaker and Erden's answer is the RoboScope system

, a low-cost robotic surgery training setup they had previously developed at Heriot-Watt. In this new work, they extend RoboScope with a custom-built haptic-enabled instrument and a real-time feedback framework. The instrument attaches to a Universal Robots UR3 collaborative arm — the kind of robot common in research labs and increasingly in light industrial settings — and is designed to be assembled and disassembled quickly, with an interchangeable tip that can accept graspers, scissors, or needle drivers

.

The core sensing hardware is an ATI Gamma six-axis force/torque (F/T) sensor — a device that measures both the magnitude and direction of forces and torques applied to it simultaneously along all three spatial axes. Crucially, this sensor is mounted at the robot's wrist, not at the instrument tip where the actual contact with tissue occurs. That location choice is deliberate. Tip-mounted sensors are fragile, expensive, difficult to sterilize, and impractical for instruments that may be replaced between procedures. A wrist-mounted sensor is more durable and easier to integrate — but it creates a signal problem. Every reading it produces is contaminated by gravity acting on the instrument itself, by the sensor's own mechanical biases, and by the instrument's changing orientation as the robot arm moves. None of those forces are contact forces. Separating the real signal from the noise is the technical heart of the problem.

The researchers had already solved that separation problem in prior work (Shaker & Erden, 2026, referencing [14]), developing a real-time compensation algorithm that removes gravitational and bias effects to isolate the true external contact force. This paper builds on that foundation by showing what to do with the clean signal once you have it.

What They Found

Once the contact force is extracted, it needs to travel a long mathematical road before the surgeon's hand can feel it. The force is measured in the sensor's coordinate frame — essentially its own local sense of "up," "forward," and "sideways." That needs to be rotated into the tool tip's frame, then into the robot's base frame, then into the haptic device's frame. Each of those transformations uses rotation matrices derived from the robot's forward kinematics, which track exactly how the arm is configured at any given moment

.

The rendering itself — translating a force number into a physical push on the operator's hand — uses a 3D Systems Touch haptic device, a consumer-grade six-degree-of-freedom pen-like controller that can push back against the user's grip with up to 3.3 N of force. The system runs at approximately 1 kHz, updating the feedback roughly every millisecond. That refresh rate matters: haptic perception is much more sensitive to timing delays than vision is. A visual lag of 50 milliseconds is barely noticeable; a haptic lag of that duration feels wrong, like a rubber glove over your fingertips.

The one remaining problem is scaling. Real tissue contact forces during surgery can swing across a wide dynamic range — from the lightest brush of a grasper to a forceful clamp. Simply passing those forces straight through to the haptic device would either saturate it (and risk damaging the hardware or hurting the user) or, if the gain is set conservatively, render fine touches imperceptible. The researchers resolve this with a nonlinear scaling function based on the hyperbolic tangent, or $\tanh$. The $\tanh$ function produces an S-shaped curve: near zero, it behaves almost linearly, preserving sensitivity to delicate forces. As forces grow large, it smoothly saturates, preventing the feedback from exceeding the device's 3 N limit. The scaled feedback force is computed as:

$${}^{\mathrm{H}}{\mathbf{f}}_{\text{scaled}}=\frac{{}^{\mathrm{H}}\mathbf{f}}{{\Vert {}^{\mathrm{H}}\mathbf{f}\Vert }_2}\tanh \left(\frac{{\Vert {}^{\mathrm{H}}\mathbf{f}\Vert }_2}{f_{\text{scale}}}\right)f_{\text{max}}$$

where $f_{\text{max}}=3$ N caps the output and $f_{\text{scale}}=7$ N (tuned empirically) controls how quickly saturation kicks in. The direction of the force is preserved by the first term; the $\tanh$ term handles only the magnitude.

The pipeline's performance is visible in Figure 6 of the paper, which plots the raw sensor signal, the compensated estimate, and the rendered haptic output across several contact events. The raw signal is noisy and drifts; the compensated signal snaps cleanly to zero during non-contact periods and rises consistently during interactions; the rendered output tracks the compensated estimate closely. The engineering is working.

To evaluate whether any of this matters to an actual human learner, the team ran a user study with 10 participants. The task was force regulation: touch a simulated gallbladder model and hold a specified contact force within ±10% of the target for at least one second. Two target force levels were tested — a "GENTLE" touch of 0.6 N (about the force of resting two small coins on your fingertip) and a "FIRM" touch of 1.2 N. Participants completed 20 randomized trials each, alternating between visual-only feedback (Haptic OFF) and visual-plus-haptic feedback (Haptic ON). No live force data was shown on a screen in either condition — the only difference was whether the haptic device pushed back.

Force Regulation Accuracy: Haptic OFF vs. ON

Label	Value
Avg. RMSE	1.35 N
Avg. Max AE	2.12 N

Source: Shaker & Erden (2026). RMSE = Root Mean Square Error; Max AE = Maximum Absolute Error. Lower is better.

The results were unambiguous. Overall task success rate doubled: 27% of trials succeeded with visual feedback alone; 54% succeeded with haptic feedback added

Task Success Rate: Haptic OFF vs. ON

Label	Value
GENTLE (0.6 N)	27 %
FIRM (1.2 N)	27 %
Overall Average	27 %

Source: Shaker & Erden (2026). Success = force held within ±10% of target for ≥1 second. Overall figures as reported; per-condition breakdown inferred from Fig. 8.

. Force regulation accuracy, measured as root mean square error (RMSE) between the actual and target force throughout each trial, improved by 36% — from an average of 1.35 N to 0.86 N. Peak force excursions — the worst-case overshoot, which is most directly associated with tissue trauma — fell by 31%, from 2.12 N to 1.46 N. Task completion time dropped by 16%, meaning participants not only regulated force more accurately but also converged on the target faster. A Wilcoxon signed-rank test — a statistical test appropriate for small samples where you cannot assume a normal distribution — confirmed that all improvements were statistically significant.

Performance Improvements with Haptic Feedback

Label	Value
Success Rate	100 %
Force RMSE (reduction)	36 %
Peak Force Error (reduction)	31 %
Task Completion Time (reduction)	16 %

100% Success Rate

Source: Shaker & Erden (2026). Success rate improvement calculated as relative change from 27% to 54%. All other metrics represent reductions.

Why This Changes Things

Put the numbers in clinical context. A 31% reduction in peak force excursions is not an abstract improvement in a training metric. In actual surgery, peak force spikes are the moments that tear tissue, snap sutures, or perforate structures. The "GENTLE" target of 0.6 N represents the kind of touch required during the most delicate phases of a procedure — the manipulation of a fragile bile duct, say, or the dissection of tissue planes near major vessels. Without haptic feedback, participants in this study regularly overshot that target by more than 2 N — more than three times the intended force. With haptic feedback, that excursion fell to roughly 1.5 N. Still imperfect, but meaningfully safer.

The 27%-to-54% doubling of success rates is also striking because it emerged after only a brief training period. Participants were not experienced surgeons; they were novices given a short familiarization session before the experiment began. Expert surgeons who have spent years compensating for the absence of haptic feedback through visual estimation and muscle memory would likely show a different — possibly smaller — improvement. The bigger opportunity is exactly where the study focuses: trainees who have not yet built those compensatory habits. Teaching them to regulate force haptically from the beginning, rather than training them to manage without it and then adding it later, may produce fundamentally better surgeons.

The cost angle is equally important. The entire hardware stack — UR3 robot arm, ATI F/T sensor, 3D Systems Touch haptic device, custom 3D-printed instrument components — is a fraction of the price of a da Vinci system, which can run to $2 million or more for acquisition alone, with annual maintenance costs in the hundreds of thousands. The RoboScope philosophy is explicit: make this technology accessible to institutions that cannot afford the commercial gold standard. That includes teaching hospitals in low- and middle-income countries, simulation centers at smaller universities, and military or remote medical training programs. If haptic surgical training can be delivered on hardware costing tens of thousands of dollars rather than millions, the pipeline of trained robotic surgeons expands dramatically.

The instrument design itself reflects a thoughtful consideration of practical deployment. The tip is interchangeable — swap out the grasper for scissors or a needle driver in seconds, without losing the force-sensing capability. The wrist-mounted sensor approach avoids the sterilization and fragility problems of tip-mounted sensors. The whole assembly threads onto the robot's adapter plate with a single rotating motion. These are not incidental engineering choices; they are what make the difference between a laboratory demonstration and something a training program can actually use.

What's Next

The study's limitations are worth naming clearly. Ten participants is a small sample, and the task — though clinically motivated — is a simplified proxy for real surgical demands. Force regulation in a clean laboratory setting, pressing on a silicone gallbladder model with one hand, is not the same as the complex, multi-degree-of-freedom, cognitively loaded environment of an actual procedure. The current system also provides only kinesthetic force feedback — the sense of being pushed back — and not cutaneous feedback, the richer texture, vibration, and slip sensations that the skin perceives. The researchers acknowledge all of this.

The most important open question may be transfer: does training with haptic feedback on RoboScope actually improve performance on real robotic surgical systems? The study shows learning within the RoboScope environment, but the field will ultimately need longitudinal studies that track trainees from the simulator to the operating room. That kind of research is harder to conduct, slower, and more expensive — but it is the evidence that would shift clinical training programs at scale.

There are also signal processing improvements to explore. The current wrist-mounted sensor approach requires careful compensation for gravity and bias — a problem the team has solved, but one that becomes more complex as the robot arm moves through unusual configurations or operates at higher speeds. Integrating machine learning approaches to force estimation, or fusing the wrist sensor data with other signals such as motor current or joint torque, could make the compensation more robust across a wider operating envelope.

What Shaker and Erden have demonstrated is a proof of concept that is compelling precisely because it is not exotic. They have taken off-the-shelf components, connected them with careful mathematics, and shown that the result meaningfully changes how well people can learn a safety-critical skill. The next generation of robotic surgeons may learn to feel before they ever touch a patient — and this work suggests that making that possible does not require a million-dollar machine. It requires the right sensor, the right algorithm, and the willingness to ask what's actually missing.