RoboPoint: a Vision-Language Model for Spatial Affordance Prediction for Robotics

Precise action guidance with image-based keypoint affordance conditioned on language instructions.

[Project Page] [Demo] [Data] [Weights]

RoboPoint: A Vision-Language Model for Spatial Affordance Prediction for Robotics [Paper]
Wentao Yuan, Jiafei Duan, Valts Blukis, Wilbert Pumacay, Ranjay Krishna, Adithya Murali, Arsalan Mousavian, Dieter Fox

Introduction

RoboPoint is a VLM that predicts image keypoint affordances given language instructions. We introduce an automatic synthetic data generation pipeline that instruction-tunes VLMs to robotic domains and needs. Compared to alternative approaches, our method requires no real-world data collection or human demonstration. In addition, RoboPoint provides a generic action space that enables language-conditioned task execution in several downstream applications such as robot navigation, manipulation, and augmented reality (AR) assistance.

Contents

• Install
• Weights
• Demo
• Data
• Train
• Evaluation

Install

./environment_setup.sh

or follow the instructions below in order.

conda create -n robopoint python=3.10 -y
conda activate robopoint

pip install --upgrade pip  # enable PEP 660 support

# this is optional if you prefer to system built-in nvcc.
conda install -c nvidia cuda=12.1 -y

pip install -e .

# this is optional if you don't need to train the model
pip install -e ".[train]"
pip install flash-attn --no-build-isolation

Model Zoo

Version	LLM	Projector	Size	Schedule
robopoint-v1-vicuna-v1.5-13b	Vicuna-v1.5	MLP2x	13B	Full FT 1 epoch
robopoint-v1-llama-2-13b	Llama-2	Linear	13B	Full FT 1 epoch
robopoint-v1-vicuna-v1.5-13b-lora	Vicuna-v1.5	MLP2x	13B	LoRA 1 epoch
robopoint-v1-llama-2-13b-lora	Llama-2	Linear	13B	LoRA 1 epoch
robopoint-v1-vicuna-v1.5-7b-lora	Vicuna-v1.5	MLP2x	7B	LoRA 1 epoch
robopoint-v1-llama-2-7b-lora	Llama-2	Linear	7B	LoRA 1 epoch

Demo

Gradio Web UI

To launch a Gradio demo locally, please run the following commands one by one. If you plan to launch multiple model workers to compare between different checkpoints, you only need to launch the controller and the web server ONCE.

Launch a controller

python -m robopoint.serve.controller --host 0.0.0.0 --port 10000

Launch a gradio web server.

python -m robopoint.serve.gradio_web_server --controller http://localhost:10000 --model-list-mode reload --share

You just launched the Gradio web interface. Now, you can open the web interface with the URL printed on the screen. You may notice that there is no model in the model list. Do not worry, as we have not launched any model worker yet. It will be automatically updated when you launch a model worker.

Launch a model worker

This is the actual worker that performs the inference on the GPU. Each worker is responsible for a single model specified in --model-path.

python -m robopoint.serve.model_worker --host 0.0.0.0 --controller http://localhost:10000 --port 20000 --worker http://localhost:20000 --model-path wentao-yuan/robopoint-v1-vicuna-v1.5-13b

Wait until the process finishes loading the model and you see "Uvicorn running on ...". Now, refresh your Gradio web UI, and you will see the model you just launched in the model list.

You can launch as many workers as you want, and compare between different model checkpoints in the same Gradio interface. Please keep the --controller the same, and modify the --port and --worker to a different port number for each worker.

python -m robopoint.serve.model_worker --host 0.0.0.0 --controller http://localhost:10000 --port <different from 20000, say 30000> --worker http://localhost:<change accordingly, i.e. 30000> --model-path <ckpt2>

If you are using an Apple device with an M1 or M2 chip, you can specify the mps device by using the --device flag: --device mps.

Launch a model worker (8-bit / 4-bit quantized)

You can launch the model worker with quantized bits, which allows you to run the inference with reduced GPU memory footprint, potentially allowing you to run on a GPU with as few as 12GB VRAM. Note that inference with quantized bits may not be as accurate as the full-precision model. Simply append --load-8bit or --load-4bit to the model worker command that you are executing. Below is an example of running with 4-bit quantization.

python -m llava.serve.model_worker --host 0.0.0.0 --controller http://localhost:10000 --port 20000 --worker http://localhost:20000 --model-path wentao-yuan/robopoint-v1-vicuna-v1.5-13b --load-8bit

Launch a model worker (multiple GPUs)

If the VRAM of your GPU is less than 24GB (e.g., RTX 3090, RTX 4090, etc.), you may try running it with multiple GPUs. Our latest code base will automatically try to use multiple GPUs if you have more than one GPU. You can specify which GPUs to use with CUDA_VISIBLE_DEVICES. Below is an example of running with the first two GPUs.

CUDA_VISIBLE_DEVICES=0,1 python -m llava.serve.model_worker --host 0.0.0.0 --controller http://localhost:10000 --port 20000 --worker http://localhost:20000 --model-path wentao-yuan/robopoint-v1-vicuna-v1.5-13b

Launch a model worker (LoRA weights, unmerged)

You can launch the model worker with LoRA weights, without merging them with the base checkpoint, to save disk space. There will be additional loading time, while the inference speed is the same as the merged checkpoints. Unmerged LoRA checkpoints are usually much smaller (less than 1GB) than the merged checkpoints (13G for 7B, and 25G for 13B).

To load unmerged LoRA weights, you simply need to pass an additional argument --model-base, which is the base LLM that is used to train the LoRA weights.

python -m llava.serve.model_worker --host 0.0.0.0 --controller http://localhost:10000 --port 20000 --worker http://localhost:20000 --model-path /home/wentaoy/checkpoints/robopoint-v1-vicuna-v1.5-13b-lora --model-base lmsys/vicuna-13b-v1.5

Visual Instruction Tuning

Download pretrained projector weights

We use pretrained projector weights from LLaVA. The projector is trained on image-text pairs from the 558K subset of the LAION-CC-SBU dataset with BLIP captions (see here). When using these projector weights, please make sure that the vision encoder and the projector type are set correctly.

For CLIP-L-336px vision encoder,

--vision_tower openai/clip-vit-large-patch14-336

For MLP-2x projector,

--mm_projector_type mlp2x_gelu

For Linear projector,

--mm_projector_type linear

Base LLM	Vision Encoder	Projection	Pretrain Data	Download
Vicuna-13B-v1.5	CLIP-L-336px	MLP-2x	LCS-558K	projector
Vicuna-7B-v1.5	CLIP-L-336px	MLP-2x	LCS-558K	projector
LLaMA-2-13B-Chat	CLIP-L-336px	Linear	LCS-558K	projector
LLaMA-2-7B-Chat	CLIP-L-336px	Linear	LCS-558K	projector

Prepare data

The data mix for instruction tuning can be found on HuggingFace at wentao-yuan/robopoint-data.

The file robopoint_1432k.json contains a list of 1432K VQA instances. An example looks like this

{
    "id": "region_ref/1033888784-63bd2a7_cam05_obj5-obj18_left",
    "image": "region_ref/1033888784-63bd2a7_cam05_obj5-obj18.png",
    "conversations": [
        {
            "from": "human",
            "value": "<image>\nIn the image, there is an item encased within a red rectangle. Pinpoint several points within the vacant space situated to the left of the object that is highlighted. Your answer should be formatted as a list of tuples, i.e. [(x1, y1), (x2, y2), ...], where each tuple contains the x and y coordinates of a point satisfying the conditions above. The coordinates should be between 0 and 1, indicating the normalized pixel locations of the points in the image."
        },
        {
            "from": "gpt",
            "value": "[(0.461, 0.527), (0.498, 0.521), (0.481, 0.521), (0.445, 0.529)]"
        }
    ]
}

The data mix consists of the following parts:

• 347K object reference data from our synthetic data pipeline
• 320K free space reference data from our synthetic data pipeline
• 150K GPT-generated instruction-following data from LLaVA
• 515K VQA data from academic-oriented tasks
• 100K object detection data from LVIS

Training

Visual instruction tuning takes around 40 hours for on 8 A100 GPUs with 80GB memory. Training scripts can be found under scripts.

If you are do not have enough GPU memory, you can reduce BATCH_PER_GPU and increase the GRAD_ACC_STEPS accordingly. Always keep the global batch size the same: NUM_NODES x NUM_GPUS x BATCH_PER_GPU x GRAD_ACC_STEPS.

Hyperparameters used in instruction tuning are provided below.

Hyperparameter	Global Batch Size	Learning rate	Epochs	Max length	Weight decay
RoboPoint-v1-13B	128	2e-5	1	2048	0

Evaluation

Where2Place, a benchmark for spatial free-space reference on challenging real world images, can be found on HuggingFace at wentao-yuan/where2place.

To evaluate on Where2Place, first run the following command to generate results

python robopoint/eval/model_vqa.py \
    --model-path wentao-yuan/robopoint-v1-vicuna-v1.5-13b \
    --image-folder datasets/where2place/images \
    --question-file datasets/where2place/point_questions.jsonl \
    --answer-file output/robopoint-v1-vicuna-v1.5-13b.jsonl

Then, run the following command to compute the accuracy

python robopoint/eval/summarize_vqa.py --answer output/robopoint-v1-vicuna-v1.5-13b.jsonl

If needed, the following command visualizes the outputs of different models together with the ground truth

python robopoint/eval/visualize_vqa.py \
    --label gpt-4o robopoint \
    --answer output/gpt-4o.jsonl output/robopoint-v1-vicuna-v1.5-13b.jsonl \
    --output output/gpt-4o-vs-robopoint \
    --num 10

Citation

If you find RoboPoint useful for your research and applications, please consider citing our paper:

@article{yuan2024robopoint,
  title={RoboPoint: A Vision-Language Model for Spatial Affordance Prediction for Robotics},
  author={Yuan, Wentao and Duan, Jiafei and Blukis, Valts and Pumacay, Wilbert and Krishna, Ranjay and Murali, Adithyavairavan and Mousavian, Arsalan and Fox, Dieter},
  journal={arXiv preprint arXiv:2406.10721},
  year={2024}
}

Acknowledgement

• LLaVA: the codebase we built upon, including the visual instruction tuning pipeline.