Skip to main content
KRAFTON’s PUBG Ally character utilizing NVIDIA ACE Text-to-Speech and behavior trees for dynamic, real-time gameplay interact

Editorial illustration for KRAFTON’s PUBG Ally uses NVIDIA ACE TTS and behavior trees for real‑time play

KRAFTON’s PUBG Ally uses NVIDIA ACE TTS and behavior...

KRAFTON’s PUBG Ally uses NVIDIA ACE TTS and behavior trees for real‑time play

2 min read

Why does this matter? Because KRAFTON is testing a new kind of in‑game teammate. PUBG Ally, the AI companion built for PUBG: BATTLEGROUNDS, entered public beta on June 17 and will run in Arcade Mode through June 30.

While most game AI relies on scripted lines, Ally leans on NVIDIA ACE’s suite of efficient models—automatic speech recognition, a 2 billion‑parameter small language model, and text‑to‑speech—to parse a player’s voice, reason about the current match and reply in real time. The result, KRAFTON calls a “co‑playable character” (CPC), a category they say sits apart from traditional NPCs by cooperating, adapting and remembering across sessions. Here’s the thing: building something that isn’t deterministic brings latency headaches and multilingual hurdles, challenges the team discussed with research lead Hyunseung Kim and project manager Yujeong Son.

But the ambition is clear: an AI squadmate that feels more like a human teammate than a pre‑written script. The upcoming beta will reveal whether the promise holds up under real‑world play.

Speech is passed to TTS and played back as Ally's voice; game actions are handed to the game side, where a behavior tree executes them and handles fast, reactive gameplay that shouldn't wait on language reasoning. All of this runs locally and continuously through the match, so the player can talk to Ally naturally while Ally also speaks and acts on its own when the situation calls for it. Why a small language model rather than a larger LLM?

What did running the Mistral-NeMo-Minitron-2B on-device give you in terms of latency, hardware reach, and player experience? For PUBG Ally, our priority was to create an AI teammate that could respond at the speed of gameplay.

Why this matters

We see KRAFTON pushing a voice‑driven AI teammate into a live shooter, using NVIDIA ACE’s small language model and local TTS pipeline. Can this model sustain fluid dialogue under fire? The system captures player speech, runs it through a 2 billion‑parameter model, then hands the resulting intent to a behavior tree that executes game actions without waiting for language processing.

All of this happens on the client, which sidesteps server latency and keeps the match fluid. For developers, the architecture suggests a path to richer, reactive companions without relying on cloud inference. Researchers may note the trade‑off between a compact model and the depth of conversational ability; the article does not reveal how nuanced the dialogue can become under stress.

Founders might wonder whether the local compute cost scales to less powerful hardware, a question the source leaves unanswered. Unclear whether this approach will generalise beyond PUBG’s specific mechanics, but the demonstration marks a concrete step toward more interactive AI characters in real‑time games.

Further Reading