---
license: apache-2.0
---
# Aligning Touch, Vision, and Language for Multimodal Perception
by Max (Letian) Fu, Gaurav Datta*, Huang Huang*, William Chung-Ho Panitch*, Jaimyn Drake*, Joseph Ortiz, Mustafa Mukadam, Mike Lambeta, Roberto Calandra, Ken Goldberg at UC Berkeley, Meta AI, and TU Dresden
[[Paper](#todo)] | [[Project Page](https://tvl.github.io/)] | [[Citation](#citation)]
This repo contains the official checkpoints for *Aligning Touch, Vision, and Language for Multimodal Perception*.
The tactile encoders comes in three different sizes: ViT-Tiny, ViT-Small, and ViT-Base, all of which are stored in
```bash
ckpt/tvl_enc
```
TVL-LLaMA, the generative counterparts, are stored in
```bash
ckpt/tvl_llama
```
## Inference
For zero-shot classification, we would require [OpenCLIP](https://github.com/mlfoundations/open_clip) with the following configuration:
```bash
CLIP_VISION_MODEL = "ViT-L-14"
CLIP_PRETRAIN_DATA = "datacomp_xl_s13b_b90k"
```
For TVL-LLaMA, please request access to the pre-trained LLaMA-2 from this [form](https://llama.meta.com/llama-downloads/). In particular, we use `llama-2-7b` as the base model. The weights here contains the trained [adapter](https://arxiv.org/abs/2309.03905), the tactile encoder, and the vision encoder for the ease of loading.
For the complete info, please take a look at the [GitHub repo](https://tvl.github.io) to see instructions on pretraining, fine-tuning, and evaluation with these models.