Imagine the sensation of tightening a bottle cap or assembling precision components: we don’t execute these tasks by measuring exact numerical values; instead, we intuitively modulate our force and direction based on the subtle changes in resistance felt through our fingertips. This instinctive perception of force is precisely what enables humans to perform delicate manipulations with such finesse.
 

However, to enable robots to acquire such perceptual capabilities, current mainstream approaches rely on sophisticated and costly six-axis force–torque sensors, which not only raise the barrier to entry but also hinder the broader adoption of this technology across diverse applications.

To address this issue, recently, RoboScience and the National University of Singapore ( NUS )’s joint research team, in their latest collaborative study “ShapeForce: Low-Cost Soft Robotic Wrist for Contact-Rich Manipulation” A low-cost, plug-and-play alternative force-torque sensor for dense-contact manipulation tasks has been developed.

Specifically, inspired by the way humans rely on the relative changes in force during contact rather than on precise force magnitudes, ShapeForce converts external forces and torques into measurable deformations of its flexible core. These deformations are then estimated via marker-based pose tracking and translated into force-like signals, enabling the robot to detect the occurrence or evolution of contact and adjust its actions accordingly. This design eliminates the need for calibration or specialized electronics to obtain exact values, instead focusing on capturing force and torque variations sufficient to support high‑density contact tasks.

Paper link: https://arxiv.org/abs/2511.19955

It is worth noting that commercially available force sensors are expensive, typically costing several thousand dollars. In contrast, ShapeForce costs only Worth about 6 dollars It can be fabricated from readily available materials while delivering performance comparable to commercial sensors. It is easily manufactured using a 3D printer and integrates perception via an RGB camera, enabling a wide range of highly tactile tasks.

Taking the task of erasing handwritten text on a shaking whiteboard—requiring precise and sustained force—as an example, a visual comparison between the pseudo‑force signal and the actual applied force reveals that ShapeForce can provide… Contact information comparable to that of a force-torque sensor 。

In addition, ShapeForce’s structural compliance further enhances its performance in… Robustness in High-Precision Tasks For example, in assembling a toy table, when the threads are misaligned, a rigid force sensor may fail to adapt even to slight offsets and often results in task failure due to difficulty in adjustment. In contrast, ShapeForce can accommodate misalignment through passive structural deformation, tolerating minute errors in high-precision tasks; its inherent fault tolerance significantly enhances the success rate of such operations.

The research team also conducted a systematic evaluation of ShapeForce in typical high‑density manipulation tasks, comparing traditional control strategies with learning‑based approaches. Whether guiding a probe through a maze with compliant motion, assisting a robotic arm in performing high‑precision USB insertion and autonomously adjusting alignment, or stably tightening a bottle cap and precisely executing plug‑in operations, the results consistently demonstrated its outstanding performance and… Generalization ability 。 

The specific tasks are shown as follows:

Maze Exploration

USB insertion

Tighten the bottle cap.

Plug-in
 

Overall, inspired by the perceptual mechanisms underlying the relative changes in human reliance on force, this study breaks the traditional trade-off between accuracy and simplicity in sensing systems and creatively develops a task‑oriented approach to force representation. Extensive experimental validation across diverse strategies and task scenarios demonstrates that ShapeForce achieves performance comparable to commercial torque sensors, significantly enhancing task success rates.

Looking ahead, RoboScience will focus on defining embodied‑intelligence hardware starting from model‑based algorithms, continuously driving the empowerment of thousands of industries through embodied intelligence.