Hexapod agent powered by Gemma4:e4b

#devchallenge #gemmachallenge #gemma

Gemma 4 Challenge: Write about Gemma 4 Submission

This is a submission for the Gemma 4 Challenge: Build With Gemma 4

Gemma4:e4b Hexapod Robot Project

What I Built

I built an AI-powered Hexapod Robot capable of autonomous navigation and dynamic gait adjustment. The project solves the complexity of coordinating 18 servos to maintain stability across uneven terrain, creating an experience where the robot can 'reason' about its movement based on sensor feedback rather than relying on hard-coded patterns.

Code

Iron-Hermes Agent Client Code

Main Repository: bradwilson331/iron-hermes
TCP Control Module: hexapod_tcp.rs
Video Handling Module: hexapod_video.rs

Server Code (Unmodified)

Hardware Server: Freenove Big Hexapod Robot Kit for Raspberry Pi

How I Used Gemma 4

I utilized the gemma4:e4b model. I chose the e4b model because it fits locally on my Mac Mini, yet provides the sophisticated spatial reasoning and logic precision required to translate high-level commands (e.g., "navigate to the red object") into explicit, coordinate-based gait adjustments and parameter generation.

System Architecture & Communication

The hexapod server operates across 2 dedicated ports:

Port 5000: Command Control (TCP)
Port 8000: Video Transmission

Operational Workflow

1. The Initial Command (Action)

Trigger: The user issues a command instructing the agent to move forward and stop when it detects an object in close proximity.
Decision: The Large Language Model (LLM) determines it is safe and time to initiate movement.
Tool Called: hexapod_tcp.rs
Execution: The agent invokes the TCP tool with the exact string protocol required by the robot's onboard hardware to initiate forward movement (e.g., CMD_MOVE#FORWARD#SPEED50). The robot physically begins walking.

2. The Sensory Polling Loop (Perception)

Because LLMs do not possess a continuous, constant "stream" of consciousness, the agent actively polls its environment at regular intervals while the robot is in motion by repeatedly executing its sensory tools:

Imaging via hexapod_video.rs: The agent executes this tool to capture the latest frame from the onboard camera (protected via an ENV_LOCK). The model analyzes the image matrix to determine whether a visual obstruction exists, such as a wall or a drop-off.
Sonar via hexapod_tcp.rs: To acquire precise, real-time distance data, the agent dispatches a telemetry request over the open TCP socket (e.g., GET_SONAR_DIST). The tool awaits the hardware return payload, receives the integer distance data (e.g., 15cm), and feeds that string back into the LLM context window for subsequent action evaluation.