Abstract
Current methods for estimating force from tactile sensor signals are either inaccurate analytic models or task-specific learned models. In this paper, we explore learning a robust model that maps tactile sensor signals to force. We specifically explore learning a mapping for the SynTouch BioTac sensor via neural networks. We propose a voxelized input feature layer for spatial signals and leverage information about the sensor surface to regularize the loss function. To learn a robust tactile force model that transfers across tasks, we generate ground truth data from three different sources: (1) the BioTac rigidly mounted to a force torque (FT) sensor, (2) a robot interacting with a ball rigidly attached to the same FT sensor, and (3) through force inference on a planar pushing task by formalizing the mechanics as a system of particles and optimizing over the object motion. A total of 140k samples were collected from the three sources. We achieve a median angular accuracy of 3.5 degrees in predicting force direction (66% improvement over the current state of the art) and a median magnitude accuracy of 0.06 N (93% improvement) on a test dataset. Additionally, we evaluate the learned force model in a force feedback grasp controller performing object lifting and gentle placement.
Introduction
Tactile sensing is a fundamental requirement for robotic systems to interact safely and effectively with their environment. However, translating raw signals from biomimetic sensors like the SynTouch BioTac into meaningful 3D force estimates remains a significant challenge. The BioTac's complex internal impedance and pressure data are difficult to model analytically, especially during shear forces. Existing learned models often overfit to specific tasks or objects. This work addresses these limitations by proposing a robust learning framework that utilizes a novel voxelized input representation and a diverse, physics-informed data collection strategy.
Methodology
The core of our approach is a deep neural network designed to capture the spatial geometry of the sensor.
Voxelized Input Representation
Instead of treating the 19 electrode readings as a flat vector, we map them into a 3D voxel grid based on their physical locations on the sensor's curved surface. This allows a 3D Convolutional Neural Network (CNN) to learn features that respect the sensor's underlying geometry, which is critical for distinguishing between normal and shear forces.
Diverse Data Collection
To ensure generalization, we collected 140,000 samples from three distinct experimental setups:
- Rigid-FT: The sensor is mounted directly to a force-torque sensor and pressed by various objects.
- Ball-FT: A robot finger interacts with a ball attached to an FT sensor, providing a wide range of contact angles.
- Planar Pushing: We infer ground truth forces by observing object motion during pushing tasks. By modeling the interaction as a system of particles with Coulomb friction, we can mathematically derive contact forces. This method is particularly effective for capturing low-magnitude forces (0.1N - 2.0N) where standard FT sensors are often noisy.
Surface-Normal Regularization
We introduce a novel loss function that weights errors based on the angle between the force vector and the sensor's surface normal. This regularization accounts for the varying reliability of tactile signals under different loading conditions.
Results
Our model demonstrates substantial improvements over both analytic baselines and previous state-of-the-art (SOTA) learned models.
Quantitative Performance
On a comprehensive test dataset, our model achieved:
- Directional Accuracy: A median angular error of 3.5 degrees, representing a 66% improvement over SOTA.
- Magnitude Accuracy: A median error of 0.06 Newtons, a 93% improvement over SOTA.
Real-World Evaluation
The learned model was integrated into a force-feedback grasp controller. In "lift and place" experiments, the robot successfully handled delicate and deformable objects, such as plastic bottles and paper cups. The accurate force feedback prevented the robot from crushing these objects or dropping them due to insufficient grip, demonstrating the model's practical utility in complex manipulation tasks.
Conclusion
This paper presents a robust method for 3D tactile force estimation. By leveraging spatial feature engineering and physics-based force inference, we achieved significant gains in accuracy and generalization. Future work will focus on mitigating sensor drift and autonomously inferring object properties to further enhance the system's autonomy in unstructured environments.
Summary updated