arXiv News

First-hand information for everyone

Site

AboutPrivacy PolicyTerms of Use

© 2026 arXiv News

PUBG Ally: a voice-enabled, embodied AI teammate that plays in real time — with caution and caveats | arXiv News

arXiv News

EnglishJapanese

Switch language

EnglishJapanese
Loading account…

First-hand information for everyone

Back to front page
Artificial IntelligenceEnglishPublished

PUBG Ally: a voice-enabled, embodied AI teammate that plays in real time — with caution and caveats

September 26, 2026arXiv: 2609.29837v1

Researchers report a new step toward game companions that can both speak with human players and act autonomously inside a fast-paced online shooter. The team behind “PUBG Ally,” described in a preprint on arXiv (arXiv:2609.29837v1), built an embodied agent for PUBG: BATTLEGROUNDS that reasons about the world, issues voice responses, and controls a game character in real time alongside human teammates.

What makes this work challenging is the need to combine two hard capabilities at once. Ally must perceive a rapidly changing game environment and respond within strict latency limits, while also interacting naturally with players and keeping its speech synchronized with its actions. To meet those requirements the authors use a layered architecture: a language-model-driven agent inspects a controlled game interface to read state, interpret player speech, keep conversational context, decide what to say, and select high-level actions. Those high-level choices then steer a faster, lower-latency control layer that executes movement, combat, and recovery in the game engine.

Because the agent’s speech and actions affect player behavior and match outcomes — and vice versa — the team emphasizes real-world data collection. They recorded nearly 39,000 gameplay sessions in which real players teamed with Ally, capturing gameplay telemetry, player speech, agent decisions, tool use, actions, and player feedback. Those records were used for iterative training and for refining how teammate quality is evaluated. In particular, the authors found gaps between offline evaluations and what players actually prefer, and they used in-game feedback and preference comparisons to update the evaluation criteria.

Practical deployment introduced further constraints. To run with low latency on players’ devices, the researchers applied model compression and context compaction; to reduce harmful or unwanted outputs, they performed targeted safety training, added runtime guardrails, and implemented memory redaction. The team also ran an international live service: players across 141 countries were surveyed, and among respondents whose play with Ally was confirmed in game records, positive survey responses exceeded negative ones by 25.1 percentage points when asked whether they would recommend Ally. Some players described Ally not merely as a tool but as a teammate or companion.

The paper is careful to flag remaining challenges. Real-time embodied assistance demands tight synchronization between speech and action, and the authors note that offline benchmarks do not always predict in-game preferences. Safety, latency, and privacy remain active engineering and research problems: compression and guardrails mitigate but do not eliminate risks, and memory redaction is one of several approaches to limit retention of player-sensitive data.

Taken together, PUBG Ally illustrates how large-language-model techniques can be integrated with fast control systems to produce voice-enabled, embodied game agents and how real-world interaction data can guide iterative improvements. The work is a clear demonstration of what is possible today, while also underlining the technical and ethical trade-offs that come with deploying conversational agents inside live, social games.

Source paper

PUBG Ally: A Conversational Embodied Agent as an AI Teammate

Artificial IntelligenceNatural Language ProcessingHuman-Computer Interaction
View on arXiv