Modality Plug-and-Play: Elastic Modality Adaptation in Multimodal LLMs for Embodied AI

Introduction

This is the official code repository for the paper "Modality Plug-and-Play: Elastic Modality Adaptation in Multimodal LLMs for Embodied AI". We present mPnP-LLM to enable fully elastic modality adaptation for LLMs via trainable latent connctions. We evaluate the performance of mPnP-LLM on a mini-split of nuScenes-QA dataset with two sensory modalities: RGB camera views and LiDAR point clouds.

Requirements

Install pytorch first and then install nuscenes-devkit with

pip install nuscenes-devkit

Install all requirements with

pip install -r requirements.txt

There might be some requirements missing in the file. Please refer to the error logs when running our code.

Creating nuScenes-QA-mini

The dataset we used in our experiments is adapted from the nuScenes-QA dataset v1.0. To create the train and validation splits for day and night scenes:

• Download nuScenes-mini split from nuScenes website

• Navigate to nuqamini folder and create path nuqamini/dataset/ and move the extracted nuScenes-mini split to it. The correct path of the dataset should look like nuqamini/dataset/v1.0-mini/data/sets/nuscenes/. Then create a path of nuqamini/dataset/v1.0-mini/data/sets/rangeprojectionoutputs/.

• Navigate to nuqamini folder and run minilidardataset_creator.py to generate range projection of the LiDAR point cloud.

• Navigate to nuqamini folder and run nuqaminidatasetcreate.ipynb. Four data splits will be created in Arrow format in the directories:

day/train/
    day/validation/
    night_80dimgaussian7/train/
    night_80dimgaussian7/validation/

Alternatively, you could download our processed dataset from huggingface. Check the dataset page here.

Prepare encoders

We use ViT-small for RGB camera views, which will be automatically downloaded when you run our training code. But we also need a pre-trained RangeViT to perceive the LiDAR inputs. Please download the pre-trained RangeViT here and put the downloaded model file under model/.

Running Modality Adaptation

Navigate to example/mpnp_llm/. We first do offline training with RGB modality on day-train split and evaluate on both day-validation split and night-validation split:

python offline_train.py

Due to low accuracy on night-split, we want to switch to LiDAR modality for better perception:

python switch_lidar.py

Alternatively, we can include both RGB and LiDAR modalities:

python add_lidar.py

Since we generate a relatively small dataset for training and validation, the obtained accuracy may have small variations due to randomness.

Citation

@article{huang2023modality,
  title={Modality Plug-and-Play: Elastic Modality Adaptation in Multimodal LLMs for Embodied AI},
  author={Huang, Kai and Yang, Boyuan and Gao, Wei},
  journal={arXiv preprint arXiv:2312.07886},
  year={2023}
}