Inspiration
In tactical games like VALORANT, good aim only gets you so far. Teams also need clear callouts, coordinated utility, and well-timed decisions. We built Comm Bot to give players a place to practice those communication skills with an AI teammate that listens and responds.
What it does
Comm Bot is a first-person tactical training sandbox where players communicate with a teammate through their microphone. Players can report enemies, request flashes, prepare coordinated swings, and tell the teammate where to play.
The bots operate with limited information. They react to delayed observations and player callouts instead of knowing every enemy’s location. This lets communication influence what the teammate believes and how it acts.
How we built it
We built the sandbox in Godot 4.4.1 using GDScript for movement, combat, utility, navigation, and bot behavior.
The voice pipeline uses Whisper for local speech recognition and Llama 3.1 through Ollama to interpret callouts. A deterministic planner and command executor turn those interpretations into physical actions. Python harnesses and headless Godot tests verify command handling, movement, perception, and round resets.
Challenges we ran into
Understanding a sentence was only part of the problem. The teammate also needed to execute it correctly: turn toward a flash destination, check visibility, navigate around obstacles, and wait when the player said “get ready.”
Keeping player instructions active was another challenge. The bot could complete a movement command and immediately return to defending the spike. We changed it to retain the requested position until the player gives another instruction.
We also had to keep bots useful without giving them access to information they should not know.
Accomplishments that we’re proud of
We connected spoken instructions to physical teammate actions using locally running models. The teammate can distinguish preparation from execution, seek a valid flash position, and navigate authored box-jump routes to gain elevation.
We also built regression coverage for these behaviors, helping us catch cases where a feature worked in isolation but failed during gameplay.
What we learned
A believable teammate needs more than language understanding. It needs consistent memory of the player’s intent, clear action priorities, and reliable physical execution.
We learned to test the entire path from communication to movement. Human playtests exposed behavior that automated checks had missed, and those failures became new regression tests.
What’s next for Comm Bot
Next, we want to improve flexible callout interpretation, expand scenario coverage, and refine combat feedback and interface clarity.
Longer term, we plan to add debriefs that explain what the teammate knew, how a callout changed its decisions, and where clearer communication could have helped.