Gravity: AI Pose & Gesture Recognition Sensor is an offline AI vision sensor that lets makers add custom static pose and hand-gesture recognition without collecting datasets or training AI models. Using the Windows software, users can quickly teach and assign IDs to up to 8 custom gestures and 8 custom full-body poses.

Once learned, actions are stored locally and the sensor can operate independently from the PC, sending recognition results to Arduino, ESP32, or micro:bit through I2C or UART. Gesture and pose recognition run as separate switchable modes, making the pose detection sensor a simple way to add personalized AI interaction to maker projects, interactive installations, and educational demos.

Add AI Pose Recognition—Without Training an AI Model

Adding custom pose recognition with a general-purpose AI vision board usually involves collecting and labeling data, training a model, converting it for the target hardware, and deploying an inference pipeline. AI Pose & Gesture Recognition Sensor already integrates the required pose-recognition capability, so makers can focus on the interaction itself instead of the AI development process.

In pose mode, the sensor detects 17 human body keypoints at approximately 10 FPS and supports up to 8 custom static poses stored locally. A learned pose can be assigned an ID and used as a trigger for lighting, displays, sound, robots, projection, or interactive installations.

Create Your Own Hand Gesture Controls

For close-range interaction, AI Pose & Gesture Recognition Sensor can learn up to 8 custom static hand gestures instead of limiting projects to a fixed gesture list. Hand mode tracks 21 keypoints with a recommended operating distance of around 0.5 m.

Custom gesture IDs can be used to start or stop a device, switch lighting effects, trigger animations, control a music project, or add personalized input to a desktop device or smart toy.

Gesture mode and full-body pose mode are separate and cannot run simultaneously; users can switch between them through the Windows software or supported communication commands.

Offline Recognition with No Cloud Dependency

Gesture and pose inference run locally on the sensor, so recognition does not require an internet connection or cloud AI service. After custom actions are learned, recognition results can be sent directly to the connected controller for local response.

This makes AI Pose & Gesture Recognition Sensor suitable for classroom demonstrations, standalone maker projects, interactive installations, smart devices, and environments where network access is unavailable or unnecessary.

Compact and Low Power for Embedded Projects

The 37 × 37 mm module is designed for compact maker devices and interactive installations. Typical operating current is approximately 40 mA at 3.3 V, while low-power mode can reduce sleep current to approximately 37 μA at 3.3 V.

A dedicated WAKEUP pin allows the main controller to place the sensor into low-power mode when recognition is not needed, making it more suitable for duty-cycled and battery-powered projects.

Built for Arduino, ESP32 and micro:bit Projects

AI Pose & Gesture Recognition Sensor provides I2C and UART communication together with a Gravity connector and 2.54 mm pin header for integration with common maker controllers. The DFRobot Arduino library provides access to recognition IDs, names, scores, bounding boxes, and hand/body keypoint data.

Arduino and ESP32 can be used directly with the available library and examples. micro:bit is supported through a compatible expansion board and MakeCode support.

For Best Recognition Results

• Use even indoor lighting and avoid strong backlight.
• Keep the hand or full body unobstructed.
• Keep the background relatively simple when possible.
• Use clearly differentiated custom gestures or poses.
• For hand interaction, keep the hand around 0.5 m from the sensor.
• For full-body pose recognition, keep the full body visible at around 2 m.
• AI Pose & Gesture Recognition Sensor is designed for static gesture and pose recognition; dynamic motion trajectories are not supported.

Features:

• Custom static pose learning: Store up to 8 custom full-body poses locally without training an AI model.
• Custom static hand gesture learning: Store up to 8 custom single-hand gestures for personalized interaction.
• 17 body keypoints: Human pose mode runs at approximately 10 FPS with a recommended distance of around 2 m.
• 21 hand keypoints: Hand gesture mode runs at approximately 5 FPS with a recommended distance of around 0.5 m.
• Switchable recognition modes: Gesture and pose recognition operate as separate modes and can be switched through the Windows software or supported commands.
• Offline local inference: Recognition runs on the sensor without cloud AI or an internet connection.
• Standalone deployment after learning: Learned actions are stored locally and can be used without keeping the PC connected.
• Low-power mode: Approximately 40 mA operating current at 3.3 V and approximately 37 μA sleep current at 3.3 V, with WAKEUP control.
• Maker-friendly interfaces: I2C, UART, Gravity PH2.0-4P, 2.54 mm header, and USB Type-C configuration.
• Development support: Arduino IDE and MakeCode support; compatible with Arduino, ESP32, and micro:bit through the appropriate connection hardware.
• Windows configuration software: Supports Windows 10/11 for live preview, custom learning, mode switching, and firmware updates.
• CE-EMC / RoHS certified.

Applications:

• Full-body pose-controlled displays and projection
• Interactive art installations
• Gesture-controlled lighting and music projects
• Smart toys and desktop interactive devices

Specifications:

• Basic Parameters
• Supply Voltage: 3.3V–5V
• Logic Level: 3.3V
• Operating Current: Typical [email protected]
• Sleep Current: Typical 37μ[email protected] (low-power mode required)
• Wake-up Method: WAKEUP pin
• Interface Parameters
• Communication: I2C / UART
• Connector: PH2.0-4P / 2.54mm pin header
• Debug/Config Port: USB Type-C receptacle
• I2C Address: 0x3A
• UART Default Baud Rate: 921600 bps
• Recognition Parameters
• Hand Keypoints: 21 keypoints
• Human Skeleton Keypoints: 17 keypoints
• Fixed Gesture Types: 13 fixed gestures
• Custom Gesture Capacity: Up to 8, deletable and re-learnable
• Custom Pose Capacity: Up to 8, deletable and re-learnable
• Gesture Recognition Range: 0.1–1m under adequate lighting (0.5m recommended)
• Pose Recognition Range: 2–4m under adequate lighting (2m recommended)
• Gesture Recognition Frame Rate: ~5 FPS
• Skeleton Recognition Frame Rate: ~10 FPS
• Hand Recognition: Single-hand only; custom gestures limited to one hand at a time, even with multiple hands present
• Pose Recognition: Multi-person supported; custom poses limited to one person at a time, even with multiple people present
• Recognition Type: Static actions only; dynamic actions not supported
• Camera Resolution: 640 × 480 (VGA)
• Camera Field of View (FOV): 89° diagonal (DFOV)
• Camera Effective Focal Length (EFL): 3.01mm
• Physical Dimensions
• PCB Size: 37mm × 37mm
• Mounting Hole Spacing: 30mm
• Mounting Hole Diameter: 3.2mm

Documentations:

• Product Wiki

Package Includes:

• 1 x Gravity: Gravity: AI Pose & Gesture Recognition Sensor
• 1 x Gravity 4-Pin Sensor Cable
• 1 x 2.54mm Black Single-Row Pin Header (10-pin)