Repository files navigation

VLM See, Robot Do (SeeDo) is a method that uses large vision models, tracking models and vision-language models to extract robot action plan from human demonstration video, specifically focusing on long horizon pick-and-place tasks. The action plan is then implemented in realworld and PyBullet simulation environment.

Note that SeeDo relies on GroundingDINO, SAM and SAM2. The code has only been tested on Ubuntu 20.04. The version of CUDA tested is 11.8, the Pytorch version is 2.3.1+cu118.

• Install SeeDo and create a new environment

gitclonehttps://github.com/ai4ce/SeeDocondacreate--nameseedopython=3.10.14condaactivateseedocdSeeDopipinstall-rrequirements.txt

• Install Pytorch (Only for CUDA 11.8 user)

pipinstalltorch==2.3.1+cu118torchvisiontorchaudio--index-urlhttps://download.pytorch.org/whl/cu118

• Install GroundingDINO, SAM and SAM2 in the same environment

gitclonehttps://github.com/IDEA-Research/GroundingDINOgitclonehttps://github.com/facebookresearch/segment-anything.gitgitclonehttps://github.com/facebookresearch/segment-anything-2.git

• Make sure these models are installed in editable packages

cdGroundingDINOpipinstall-e .

And do the same with segment-anything, segment-anything-2

• We have slightly modified the GroundingDINO

InGroundingDINO/groundingdino/util/inference.py, we add a function to help inference on an array of images. Please paste the following function intoinference.py.

defload_image_from_array(image_array:np.array)->Tuple[np.array,torch.Tensor]:transform=T.Compose(        [T.RandomResize([800],max_size=1333),T.ToTensor(),T.Normalize([0.485,0.456,0.406], [0.229,0.224,0.225]),        ]    )image_source=Image.fromarray(image_array)image_transformed,_=transform(image_source,None)returnimage_array,image_transformed

• The code still uses one checkpoint from segment-anything.

Make sure you download it in the SeeDo folder.default orvit_h:ViT-H SAM model.

• Obtain an OpenAI API key and create akey.py file underVLM_CaP/src

cdVLM_CaP/srctouchkey.pyecho'projectkey = "YOUR_OPENAI_API_KEY"'>key.py

There are mainly four parts of SeeDo. To ensure the video is successfully processed in subsequent steps, useconvert_video.py to convert the video to the appropriate encoding before inputting it. Theconvert_video.py script accepts two parameters:--input and--output, which specify the path of your original video and the path of the converted video, respectively.

1. Keyframe Selection Module

   get_frame_by_hands.py: Theget_frame_by_hands.py script allows selecting key frames by tracking hand movements. It accepts two parameters.

   --video_path, which specifies the path of the input video.

   --output_dir, which designates the directory where the key frames will be saved. Ifoutput_dir is not specified, the keyframes will be saved to./output by default. For debugging purpose, the hand image and hand speed plot will also be saved in this directory.

2. Visual Perception Module

   track_objects.py: Thetrack_objects.py script is used to track each object and add a visual prompt for the objects. It also returns a string containing the center coordinates of each object in the key frames. The script accepts three parameters.

   --input is the video converted to the appropriate format.

   --output specifies the output path for the video with the visual prompts.

   --key_frames is the list of key frame indices obtained fromget_frames_by_hands.py.

   This module will return abox_list string stored for useage in VLM Reasoning Module

3. VLM Reasoning Module

   vlm.py: Thevlm.py script performs reasoning on the key frames and generates an action list for the video. It accepts three parameters.

   --input is the video with visual prompts added by the Visual Perception Module.

   --list is the keyframe index list obtained from the Keyframe Selection Module.

   --bbx_list is thebox_list string obtained from the Visual Perception Module.

   This module will return two strings:obj_list representing for the objects in the environment;action_list representing for the action list performed on these objects.

4. Robot Manipulation Module

   simulation.py: Thesimulation.py script accepts three parameters:obj_list,action_list,output. It first initializes a random simulation scene based on theobj_list, and then executes pick-and-place tasks according to theaction_list, and finally write the video to output.

   Example usage:python simulation.py --action_list "put chili on bowl and then put eggplant on glass" --obj_list chili carrot eggplant bowl glass --output demo2.mp4

   Note that this part uses a modified version of the Code as Policies framework, and its successful execution depends heavily on whether the objects are already modeled and whether the corresponding execution functions for actions are present in the prompt. We provide a series of new object models and prompts that are compatible with our defined action list. If you want to operate on unseen objects, you will need to provide the corresponding object modeling, and modify the LMP and prompt file accordingly.

   We provide some simple object modelings of vegetables on hugging face. Download fromhttps://huggingface.co/datasets/ai4ce/SeeDo/tree/main/SeeDo

   There will be anassets.zip file, extract that file intoassets and make sure this folder is under the path of VLM_CaP.VLM_CaP/assets will then be used bysimulation.py for simulation.

   It will write out a video of robot manipulation of a series of pick-and-place tasks in simulation.