Introduction
In robotic manipulation, detecting "slip" has traditionally been treated as a binary classification problem: either an object is slipping, or it is not. However, a simple yes/no response is often insufficient for delicate or high-precision tasks. If a robot only knows that a slip is occurring without understanding its magnitude, it might apply excessive force—potentially crushing a fragile object—or fail to apply enough force to stop a rapid slide.
This research introduces a nuanced, tactile-only framework that moves beyond binary detection to quantify slip severity (relative velocity). By measuring the velocity between the gripper and the object, the system enables sophisticated feedback control that can adjust grip dynamically without overcompensating.
Methodology: Vector Field Analysis
The core of this approach utilizes a GelSight Mini tactile sensor. Unlike traditional methods that process raw images, this framework analyzes the deformation vector fields created by the displacement of black markers on the sensor's flexible gel surface.
The researchers extract several physics-informed features from these marker movements:
- Net Velocity ($\bar{v}_{net}$): The overall speed of gel deformation.
- Divergence ($\nabla \cdot \vec{D}$): Measures expansion or contraction, which is closely linked to changes in normal force.
- Curl ($\nabla \times \vec{D}$): Measures rotational shifts, capturing torsional stress during manipulation.
- Contact Area ($A_n$): Reconstructed in 3D to determine the extent of the object's contact with the sensor.
These features are fed into a dual-model pipeline. A Random Forest (RF) or Gradient Boosting (GB) classifier handles slip detection, while a Long Short-Term Memory (LSTM) network estimates the slipping velocity (cm/s). The LSTM was specifically chosen for its ability to capture the temporal, time-series nature of deformation events.
Experimental Setup and Datasets
The framework was validated using a UR5e robotic arm across a variety of objects. The training dataset included 15 objects of diverse shapes and textures, categorized into three scenarios: static grasping, active grasping (increasing pressure), and deliberate slipping.
To test generalization, the system was evaluated on five "unseen" objects not present in the training set, including a porous sponge, a wooden plank, and a pair of scissors. Ground truth for slipping velocity was established using external ArUco markers and an Intel RealSense camera to ensure the models were learning the correct speeds.
Key Insights and Results
A major finding of this work is the effectiveness of physics-informed features. By using vector calculus (divergence and curl) rather than raw pixels, the models become more robust to variations in object geometry and lighting.
The system achieved an average detection accuracy of 92% on unseen objects. In terms of severity estimation, the LSTM model reached a Mean Absolute Error (MAE) of 0.26 cm/s in cross-validation. Notably, the system performed well even on challenging objects like a porous sponge, where inconsistent deformation typically leads to noisy data.
In closed-loop control tests, the robot successfully stabilized a smooth PVC pipe—a low-friction object—by detecting the slip and adjusting its gripper position in real-time to bring the slipping velocity to zero without any prior knowledge of the pipe's weight or texture.
Conclusion and Future Work
This framework demonstrates that tactile deformation fields provide sufficient information for both high-accuracy slip detection and precise velocity estimation. By moving beyond binary feedback, robots can achieve more human-like dexterity in handling fragile or unpredictable objects.
Future research will focus on expanding the framework to estimate angular (rotational) slipping velocities and integrating adaptive force control for even more delicate manipulation tasks, such as complex industrial assembly.
Summary updated