git clone https://github.com/tyb197/RoboAfford.git
cd RoboAfford
conda create -n roboafford python=3.10 -y
conda activate roboafford
pip install -r requirements.txt

The above environment supports evaluation of both closed-source models (GPT-4o, Claude-3.5-Sonnet, Gemini-1.5-Flash, and Gemini-1.5-Pro) and the following open-source models: Qwen2-VL and Qwen2.5-VL.

For other open-source models, we recommend creating separate conda environments for each model following their official GitHub repositories.

For closed-source model evaluation, configure your OpenAI API key in eval_closed_source_models.sh:

export OPENAI_API_KEY="Your OPENAI_API_KEY"

Then run the following command:

bash eval_closed_source_models.sh

For open-source model evaluation, we recommend downloading the checkpoints from Hugging Face first, then loading them from local paths. Supported models include:

• Molmo: allenai/Molmo-7B-D-0924
• Qwen2-VL: Qwen/Qwen2-VL-7B-Instruct
• LLaVA-Next: lmms-lab/llama3-llava-next-8b
• SpaceMantis: remyxai/SpaceMantis
• RoboPoint: wentao-yuan/robopoint-v1-vicuna-v1.5-13b
• Qwen2.5-VL: Qwen/Qwen2.5-VL-3B-Instruct, Qwen/Qwen2.5-VL-7B-Instruct

Run the evaluation with:

python inference.py \
--model qwen25vl-7b \
--model_path /your_path_to_checkpoint/Qwen2.5-VL-7B-Instruct \
--annotations "annotations_absxy.json" \
--output "./results_open_source_models/qwen25vl-7b.json"

We provide annotation files in two formats to accommodate different model output formats:

• annotations_normxy.json (values between 0-1)
• annotations_absxy.json (pixel values)

You can specify your preferred format using the --annotations argument when running the evaluation. For implementation details, please refer to the eval_open_source_models.sh script.

After evaluation, visualize the results using:

python visualize.py --json_path "./results_open_source_models/qwen25vl-7b.json" --image_dir "./images" --output_dir "./visualize_results"