LingBot-Touch-Tiny

LingBot-Touch-Tiny

Encodes tactile sensor signals into latent features for robot manipulation research, covering magnetic, piezoresistive and visuotactile sensing. The repo holds the weights and config only.

Made by robbyant on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: LingBot-Touch-Tiny is the encoder of LingBot-Touch, a tactile representation model for robotics built on a JEPA-style design. It learns features of contact structure and how contact changes over time, by predicting in latent space. It covers three kinds of touch sensing: magnetic, piezoresistive and visuotactile.

Who it suits: Robotics researchers who work with touch sensors and dexterous hands and want a pretrained encoder for tactile signals.

Industry | Who uses it | How they use it | Example
Robotics | Manipulation researchers | Turn tactile signals into features for grasping tasks | Feed fingertip readings into a downstream policy
Sensor research | Hardware teams | Compare different tactile sensor types in one space | Map magnetic and piezoresistive data to the same format
Academia | Students | Study latent prediction for tactile data | Probe the encoder's output features

If you want a ready-to-run robot controller, you can skip it, since the repo holds only weights and config.

Where to use it:

• Research code that loads an encoder and a config file
• The separately distributed LingBot-Touch inference implementation
• Windows, macOS or Linux, for downloading the files

Use case | Who | What they get back
Tactile features | Researchers | Token-level latent representations of touch
Manipulation research | Robotics teams | Features for downstream perception and control
Sensor comparison | Hardware teams | One representation across three sensing types

How to set it up:

• Python with pip
• The LingBot-Touch inference implementation, which is distributed apart from this repo
• A sensor adapter of your own, because the repo has none

1. Install the Hugging Face hub package.
2. Download the repo, which gives you encoder.pt and config.json.
3. Keep the two files together and load them with the inference implementation.

pip install -U huggingface_hub
hf download robbyant/lingbot-touch-tiny

How to use it: Pass the encoder normalized tactile signals together with taxel coordinates and a choice of sensing type. It returns latent features that you read out for your task. The card recommends querying them with attention.

Tool | What it does
encoder.pt | The frozen target encoder's weights
config.json | The inference architecture and sensing-type routes

Example: You normalize the readings from a robot hand's five fingers and select the magnetic sensing type. The encoder returns token-level latent features with their time, finger and position labels, and you feed them to a task head.

For newcomers and experts: If you are new, this is a building block for robotics research, not an app. Read about the input shape and the sensor setup before you try it.

The input has shape batch, time, 5 fingers, points and 3 values. Fingers always run thumb, index, middle, ring and pinky. X and Y are normalized to between minus one and one and Z to between zero and one. For piezoresistive and visuotactile sensors, the single value is placed in Z and both X and Y are zeroed.

Tactile signals are normalized
Taxel coordinates and sensing type are chosen
The shared Transformer encodes them
Latent tokens come out with addresses

Good to know:

• Sensor geometry, handedness, point order and measurement ranges must be adapted before inference
• Raw RGB images are not inputs for the visuotactile type
• License: Apache 2.0
• Cost to run: none named
• Data it reaches: the tactile signals you supply and the files downloaded from Hugging Face
• Not for: use as a full toolkit, since it has no inference code, sensor adapters, sample data, predictors, task heads or training logs

On Hugging Face, 13 November 2026

• Task: robotics
• Downloads in the last 30 days: 380
• Likes: 13
• License: apache-2.0
• Tags: tactile, robotics, representation-learning

Links

• Model: https://huggingface.co/robbyant/lingbot-touch-tiny
• Files: https://huggingface.co/robbyant/lingbot-touch-tiny/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download robbyant/lingbot-touch-tiny

Then follow the usage in its model card. Nothing in this kit runs by itself.