GenTact-Prox: Design, Mapping, and Contact Anticipation with 3D-printed Whole-Body Tactile and Proximity Sensors

Skin units are procedurally generated using Blender geometry nodes, automatically conforming to robot morphology through four sequential operations: dermis molding, sensor distribution, routing, and wiring. A designer only needs to provide the robot model and tune design parameters—no manual modeling required. The skin's flush fit allows it to snap directly onto a robot's body without adhesives, and each unit is fabricated using standard multi-material FDM 3D printing with rigid PLA and conductive PLA filaments. Capacitive electrodes are embedded inside the dermis layer (rather than exposed on the surface) to prevent signal saturation at direct contact, allowing the sensors to remain sensitive to the weaker proximity signals from nearby objects.

Each skin unit senses touch and proximity using self-capacitance. Every 3D-printed conductive electrode creates a small electrostatic field around itself. When a person's hand, a tool, or any nearby object enters that field, it distorts it—and the sensor picks up the change. The closer the object, the larger the disturbance. This lets the robot "feel" objects approaching before they make contact.

Distance is estimated by measuring how quickly the sensor's circuit charges and discharges—no contact required. The sensors are read by an ESP32-C6 microcontroller at 15–30 Hz, providing a continuous stream of proximity data across the robot's body. Each sensor's detection range depends partly on its surface area; in our evaluation, ranges spanned from 2.6 cm to 18 cm.

Knowing that a sensor can detect something isn't enough—you also need to know where it can detect reliably. We introduce the concept of Perisensory Space (PSS): the body-centric volume of space around the robot where the skin's sensors can deliver useful proximity information. Unlike structured sensors (e.g., cameras or time-of-flight sensors) with well-defined fields of view, our 3D-printed capacitive sensors are irregularly shaped and distributed, making it impossible to calculate the PSS analytically.

Instead, we take a data-driven approach: a robot sweeps an object over each skin unit while recording sensor readings and object positions. An ensemble of 100 neural networks is trained on this data to predict where a nearby object is located. When multiple networks in the ensemble disagree on the answer, that disagreement signals an unreliable region—outside the effective PSS. We show that this disagreement (uncertainty) is strongly correlated with true localization error, making it a reliable filter. Regions within 0–8 cm of the sensor consistently fell within the reliable PSS, suitable for contact anticipation.

We designed, fabricated, and evaluated five unique skin units for the Franka Research 3 (FR3) robot arm, covering the entire arm from base to wrist. Key results:

• Detection Range: Sensors detected objects from 2.6 to 18 cm away, with larger electrode surface areas producing longer ranges.
• PSS Mapping: The ensemble's uncertainty was strongly correlated with true localization error—when the model was confident, it was accurate. Regions within 0–8 cm of the skin were consistently reliable for contact anticipation.
• Collision Avoidance: The robot successfully detected a human hand approaching its path and slowed down and avoided the obstructed region before physical contact occurred, then resumed normal motion once the path was clear.
• Low Cost and Open Source: Each skin unit costs less than $25 USD to fabricate. All procedural design operations and printable files are open-sourced.