The universal neural representation protocol for robotic tactile sensing. High-fidelity. Latency-critical. Hardware-agnostic
Discrete taxel readings are spatially aliased, sensor-specific, and latency-bound. Neural Fields encode touch as a continuous, differentiable function over a learned representation space — enabling interpolation, generalization, and cross-hardware semantic equivalence.
Traditional tactile pipelines treat each sensor as a proprietary, hardware-bound array of pressure values. The result: non-transferable representations, lossy spatial resolution, and inference pipelines locked to specific hardware generations.
TactileFields encodes contact geometry and force distribution as a continuous implicit neural function. Query any point in the contact manifold at arbitrary resolution. Differentiable end-to-end. Portable across hardware generations.
// Instantiate a TactileField from raw sensor readings
import tactilefields as tf
field = tf.TactileField.from_sensor(
sensor_type="DIGIT_v2",
raw_frame=sensor.read(), // any hardware
resolution=FieldResolution.HIGH_FIDELITY,
latency_budget_ms=0.8
)
// Query continuous contact pressure at arbitrary spatial coordinates
contact = field.query(x=0.42, y=0.17, derivative_order=2)
// → ContactEvent(pressure=0.73, shear=(0.02, -0.11), curvature=...)
The TactileFields schema is a universal abstraction over physical sensor modalities. Certified adapters exist for vision-based tactile sensors, resistive arrays, piezoelectric films, and capacitive grids — with new hardware partners added quarterly.
Hardware adapters execute in a dedicated low-latency thread with strict 0.8ms budgets, SIMD-accelerated preprocessing, and zero-copy field construction via shared memory.
Models trained against the TactileFields schema deploy unchanged across hardware generations. Replacing a sensor requires no retraining — only adapter recertification.
All certified adapters expose a differentiable forward pass, enabling end-to-end gradient flow from field query through policy network for imitation and RL pipelines.
| Capability | Specification | Status | Notes |
|---|---|---|---|
| Vision-based tactile encoding | 512×512 field, fp16 | Ready | GelSight, DIGIT, Tactip certified |
| Resistive array encoding | 4096 taxel, linear interp. | Ready | Weiss, RoboSkin certified |
| Multi-fingered hand fields | Per-segment manifold | In Progress | Optimus, LEAP hand |
| Streaming field compression | 94.7% lossless, 12kbps | Ready | For distributed inference |
| Proprioceptive fusion | Joint-state + tactile | In Progress | Q2 2025 |
| Cross-modal interpolation | Thermal + pressure + shear | Planned | v1.0 milestone |
The TactileFields specification is developed in collaboration with research groups whose sensor hardware and learning frameworks constitute the state of the art in robotic tactile perception.
The TactileFields SDK v0.9.2-alpha is available to qualified research groups and robotics engineering teams. Access includes the core field encoding library, certified hardware adapters, reference policy implementations, and the full specification draft.