Physical AI: Adding Intelligence to Existing Robots

BeDraft proposes Physical AI, an approach that adds sensors and edge computing modules to already deployed robots to grant them natural language understanding and environmental awareness capabilities, instead of building new robots.

5-Layer System Architecture

• Input Layer: Data collection via microphones, cameras, LiDAR, etc.
• Perception Layer: Performs VLM, speech recognition, and object and spatial recognition
• Localization Layer: Tracks current location using maps and LiDAR
• Decision Layer: Interprets natural language commands and selects actions
• Control Layer: Converts AI outputs into robot movements via a Control Harness

Key Technical Differentiators

• Control Harness: Transforms the probabilistic outputs of VLMs into deterministic robot actions. Instead of directly passing sentences generated by AI, it maps them to an allowed Action Set to ensure safety.
• Platform Independence: Since it does not modify the robot body's firmware, it separates the AI layer from the robot control layer, allowing easy scaling across various robot platforms.
• Edge Resource Management: It is crucial to optimize latency, memory, heat, and power consumption between VLM inference, sensor processing, and control loops on edge devices such as the NVIDIA Jetson Nano.

Key Demonstration Cases

• Boston Dynamics Spot: Built a system by adding an external on-device module to recognize Korean voice commands and perform actions.
• Low-cost Quadruped Robot: Equipped with Jetson Nano and LiDAR to transition from traditional remote control methods to a destination-based autonomous walking system.