Skip to content

Latest commit

 

History

History
164 lines (132 loc) · 10.2 KB

README.md

File metadata and controls

164 lines (132 loc) · 10.2 KB
Raw

Robot Commander

Requirements

Local models

This project is currently conceived for inference usage only. If you want to fine-tune any of the models, you will have to research and implement the required steps yourself.

Local model support will be relying heavily on llama.cpp, whisper.cpp, and their derivatives when possible. The comon interface of this project is created in python, however, to also allow easy integration and testing of native pytorch models and ML libraries.

External API

This project currently supports requests to public OpenAI and Anthropic API servers. To be able to access their APIs you must have an account at OpenAI, and/or Anthropic with a valid payment method set.

Results

The following three example scenarios showcase preliminary results. The commander always first waits for a fully generated audio response, and only then proceeds to generate a valid goal pose command for Navigation2. All examples were executed with locally hosted models. The used robotic platform was developed at the BUT Robotics and AI Group.

  1. Sending the robot to an unspecified location (voice prompt, speech response):

    explore_onethird_resolution_amplified_speech.mp4

  2. After the pose contexts are updated with the new location, sending the robot back to the origin pose (voice prompt, speech response):

    return_onethird_resolution_amplified_speech.mp4

  3. Sending the robot to the newly acquired hall pose (voice prompt, speech response):

    hall_onethird_resolution_amplified_speech.mp4

Shortcomings and TODOs

  1. The current design works best when translating prompts with defined keywords into exact pose messages. Some RAG-like feedback context about the current state could be included to allow for handling more complex prompts. This is currently a work-in-progress.
  2. The ROS interface currently only supports generating PoseStamped commands. Integration of other message types and some decision-making for more complex behaviors might be useful, but it will not be a high priority for now.
  3. The state of quantization for TTS is currently not fully developed. This will have to be managed by the external project owners, unless we want to do the contributions ourselves.

Setup

Build

  1. Install system dependencies: 
    sudo apt install ninja-build portaudio19-dev nlohmann-json3-dev

    NOTE: Also make sure you have installed ROS and the following dependencies: python3-colcon-common-extensions, python3-rosdep, and python3-vcstool.

  2. Download this repo and update ROS dependencies: 
    git clone https://github.com/Imaniac230/robot-commander.git && . /opt/ros/<ros-distro>/setup.bash && rosdep install --from-paths robot-commander/ --ignore-src -r -y
  3. This project relies on the independent robot_commander_library package located in library_vendor/. The other external projects listed in library_vendor/libraries.repos are required for usage with local models only (if you're only going to make requests to external servers, you only need robot_commander_library). The input package is also optional, and provides a gamepad interface with support for additional interactions with the DualSense PS5 controller. To download all the external libraries use the library_vendor_setup.sh script (or run the included commands manually): 
    cd robot-commander/ && ./library_vendor_setup.sh
  4. Build the packages: 
     colcon build

    NOTE: If you want to build the .cpp libraries with CUDA support enable the USE_CUDA option:

    colcon build --cmake-args -DUSE_CUDA=ON

    NOTE: If you want to compile llama.cpp and whisper.cpp with different options than defined in this project, please refer to their respective instructions for more detailed compilation steps.

  5. Source the current ROS workspace: 
    . install/setup.bash

Local models

Download and quantize

This project requires all models to be downloaded and/or quantized manually before running any examples. To get the most comprehensive and up-to-date instructions, please always follow the instructions in the llama.cpp and whisper.cpp README. Here are basic simplified example steps to prepare the whisper, llama3, and bark models for use:

Whisper

  1. Download the official whisper repo from OpenAI: 
    git clone https://github.com/openai/whisper.git
  2. Download one of the official models from OpenAI, ex: 
    wget https://openaipublic.azureedge.net/main/whisper/models/e5b1a55b89c1367dacf97e3e19bfd829a01529dbfdeefa8caeb59b3f1b81dadb/large-v3.pt

    Or download from OpenAI on huggingface, ex:

    git clone https://huggingface.co/openai/whisper-large-v3

    Make sure that you have git lfs installed: git lfs install

  3. Use the convert-pt-to-ggml.py script from whisper.cpp to convert, ex: 
    python3 whisper-convert-pt-to-ggml.py <path-to-downloaded-model-file>/large-v3.pt <path-to-official-whisper-repo> <path-to-converted-file> && mv <path-to-converted-file>/ggml-model.bin <path-to-converted-file>/ggml-model-f16-large-v3.bin

    To convert the model downloaded from huggingface, use convert-h5-to-ggml.py instead, ex:

    python3 whisper-convert-h5-to-ggml.py <path-to-downloaded-model-files> <path-to-official-whisper-repo> <path-to-converted-file> && mv <path-to-converted-file>/ggml-model.bin <path-to-converted-file>/ggml-model-f16-large-v3.bin
  4. Quantize the converted model, ex.: 
    whisper-quantize <path-to-converted-file>/ggml-model-f16-large-v3.bin <path-to-quantized-file>/ggml-model-q4_0-large-v3.bin q4_0

Llama3

  1. Download one of the official models from Meta on huggingface, ex.: 
    git clone https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct

    Make sure that you have git lfs installed: git lfs install

  2. Use the convert_hf_to_gguf.py script from llama.cpp to convert, ex.: 
    python3 llama-convert_hf_to_gguf.py <path-to-downloaded-model-files> --outtype f16
  3. Quantize the converted model, ex.: 
    llama-quantize <path-to-converted-file>/ggml-model-f16.gguf <path-to-quantized-file>/ggml-model-q4_0.gguf Q4_0

Bark

  1. Download the official model files from SunoAI on huggingface, ex: 
    git clone https://huggingface.co/suno/bark

    Make sure that you have git lfs installed: git lfs install

NOTE: The GGML format and quantization for bark are currently only experimental. Use the full pytorch models if you want the best results.

  1. Use the convert.py script from bark.cpp to convert, ex.: 
    python3 bark-convert.py --dir-model <path-to-downloaded-model-files> --use-f16 && mv <path-to-converted-file>/ggml_weights.bin <path-to-converted-file>/ggml-model-f16.bin
  2. Quantize the converted model, ex.: 
    bark-quantize <path-to-converted-file>/ggml-model-f16.bin <path-to-quantized-file>/ggml-model-q4_0.bin q4_0

Usage

  1. Specify your commander parameters in params/commander_params.yaml.

  2. Launch the ROS commanders:

    SUNO_OFFLOAD_CPU=True SUNO_USE_SMALL_MODELS=True ros2 launch robot_commander_py commanders.launch.py

    If you have enough memory to hold the full Bark models, you can disable the SUNO_USE_SMALL_MODELS option. If you want to use CUDA support with Bark, you can disable the SUNO_OFFLOAD_CPU option. Please refer to the Bark README for more details.

    NOTE: Local quantized Bark TTS agent server is currently only experimental. For best results, use the raw pytorch models, which do not support a server mode, and must be loaded dynamically for each request. To use the TTS server host, disable the use_pytorch parameter for the chat commander text_to_speech section.

  3. Start providing voice commands in natural language using a "push-to-talk" interface exposed through the /record_prompt topic.

    NOTE: If you want to use the included gamepad interface node, use the launch_with_gamepad:=True argument.

  4. To update the pose "keyword" mappings with the current robot position call the update_pose_context ROS service:

    ros2 service call /update_pose_context robot_commander_interfaces/srv/UpdateContext "{"keyword": "home"}"

  5. To store the current pose "keyword" mappings into the specified file call the save_contexts ROS service:

    ros2 service call /save_contexts std_srvs/srv/Trigger

Local model servers

To launch the locally hosted agent servers:

  1. Specify your agent parameters in params/agent_params.yaml.
  2. Launch the local agent servers: 
    ros2 launch robot_commander_py agents.launch.py

Other examples

Some other examples, such as the voice_prompted_image.py script use a wrapper around the openai python library. If you want to use that, you will also have to install the library: pip install openai.