GitHub radar
One tactile model for many sensors and robot hands
LingBot-Touch turns readings from different tactile sensors into one shared format any robot hand can use — no separate model per sensor type.
It's a model that turns tactile sensor signals (magnetic, piezoresistive, visuotactile) into one shared representation.
When it helps
Useful if you're building a tactile-sensing robot hand or researching skill transfer across sensors; without a tactile sensor or robot hand there's little you can do.
Pros
- One model instead of a separate one per sensor type
- Transfers to new hardware without extra training
- Runs on a regular laptop, no GPU needed
- Quality improves as training data grows
Cons
- Weights require a separate request on Hugging Face or ModelScope
- Repo only ships the inference part, not training
- Without a tactile sensor you can only run sample data
How to set it up — step by step
- 1Open the repo https://github.com/Robbyant/lingbot-touch and read the README.
- 2Send the agent the text from the block below.
- 3Once you get access to the Tiny model weights on Hugging Face, let the agent download them.
- 4Check that the agent set up the environment (Python 3.12, PyTorch) and ran inference on the Sharpa samples.
- 5Open the resulting features file and confirm it contains a latent tensor with representations.
Text for your agent
Copy this and send it to your agent — Claude Code, Codex, any of them:
Set up LingBot-Touch from https://github.com/Robbyant/lingbot-touch — follow the README to set up the environment (Python 3.12, PyTorch), ask me for access to the Tiny model weights on Hugging Face, run inference on the Sharpa sample data included in the repo, and show me the resulting feature file.
The model itself is light and runs fine on a CPU, no graphics card required. But it is only practically useful with an actual tactile sensor or robot hand — without one, all you can do is run the sample recordings included in the repo.
Open on GitHub▌ More finds