Vision-Guided Manipulation with Gemini & Robot Arms

This repository contains a demonstration of a Vision-Language-Action (VLA) system using the SO-101 robot arm, an attached USB camera, and the Gemini Robotics ER 1.5 model for zero-shot object detection and pointing.

1. Prerequisites

Before running the script, set up your environment and install dependencies.

Environment setup (Linux/macOS)

The recommended way to set up your environment is based on the LeRobot installation instructions.

• Install Conda (Miniforge recommended):

wget "https://github.com/conda-forge/miniforge/releases/latest/download/Miniforge3-$(uname)-$(uname -m).sh"
    bash Miniforge3-$(uname)-$(uname -m).sh

• Create and Activate Environment:

conda create -y -n gemini-robotics-pointing python=3.10
    conda activate gemini-robotics-pointing

• Install ffmpeg:

conda install ffmpeg -c conda-forge

• Install Python dependencies:

pip install opencv-python numpy scipy pillow google-genai \
    "lerobot[feetech]" mujoco urchin placo requests

2. Get configuration parameters and run the script

To identify the USB port for your robot arm, use the lerobot-find-port command.

To identify the correct camera index, use the lerobot-find-cameras opencv command.

Provide configuration parameters as command-line arguments.

Execution

Before running the script, manually position the arm so it does not block the camera's view of the ChArUco board.

To run the script, provide your API key and adjust the other required values as needed:

python workshop.py \
--api-key "MY_GEMINI_API_KEY" \
--port "/dev/tty.usbmodem12345" \
--robot-id "my_so101_follower" \
--camera-index 0

Need a Gemini API key? Go to Google AI Studio to get one!

To use an existing arm calibration file, include the arg --calibration-dir:

python workshop.py \
--calibration-dir "path/to/calib_dir" \
...rest of args...

This should point to the directory that contains the calibration file, not the file itself. The filename must match the robot-id, e.g. my_so101_follower.json.

Calibration step

The first time you run the script, or if you use the --recalibrate flag, it will perform the ChArUco calibration.

• Ensure the ChArUco board is in the camera's view and unobstructed before

running the script.

• The script will display a window showing the detected board (if successful)

and save the homography matrix to homography_calibration.npy. **Press q to close the window and continue the script.**

• The workshop's ChArUco board includes a registration outline which sets a

known physical X (forward) and Y (left) distance from the robot's base to the Anchor Corner (ID 0) of the board. If you're placing the board differently, measure and provide the actual distance as the --board-origin X Y argument.

Interactive loop

After successful calibration, the script will enter an interactive loop:

• The robot moves to the home position.
• The script prompts: "⌨️ What should I point at? (e.g., 'blue block',

'pen'):"

• The camera captures an image, and the image is sent to Gemini with your

prompt.

• Gemini returns the 2D pixel coordinate of the object's center.
• The script uses the saved Homography matrix to convert the pixel coordinate

to real-world (X, Y) robot coordinates.

• The robot executes a sequence: Home -> Hover 10cm above table -> Descend to

Point 2cm above table.

• The loop repeats until you type q.

Key script arguments

Argument	Description	Default Value	Required?
--api-key	Your Google AI Studio API Key.	N/A	Yes
--port	Serial port connected to the robot arm.	N/A	Yes
--robot-id	ID of the robot arm; must match calibration filename without extension.	N/A	Yes
--camera-index	Index of the USB camera (try 0, 1, or 2).	N/A	Yes
--calibration-dir	Directory containing the arm calibration files, when not using the default location or the lerobot-calibrate command.	N/A	No
--backend	Kinematics solver: lerobot, argo, or mujoco.	argo	No
--board-origin	X (forward) and Y (left) robot coordinates (meters) of the ChArUco board's Anchor Corner (ID 0).	0.29 0.0525	No
--recalibrate	Flag to force recalibration.	N/A	No