AI Papers Reader

Personalized digests of latest AI research

View on GitHub

Voice-Enabled AI Teammate Plays PUBG Alongside Humans

In team battle-royale games, a useful AI partner must do more than aim and shoot: it must understand spoken commands, call out threats, and keep words synchronized with actions under strict latency. The PUBG Ally Team introduced PUBG Ally, a voice-enabled embodied agent deployed as an AI duo partner in PUBG: BATTLEGROUNDS. The paper describes a live-service beta in which Ally reasons, acts autonomously, and communicates with a human teammate on the Sanhok map. Such agents must combine real-time gameplay with conversational interaction, two capabilities that are difficult to integrate under one control loop.

Ally separates deliberation from fast control. A small language model acts as System 2, invoked by selected events rather than a fixed clock. Through a bounded tool interface, it inspects relevant game state, retrieves knowledge and memory, interprets player speech, decides what to say, and chooses high-level actions. A deterministic behavior tree acts as System 1, running every game tick to translate those choices into movement, combat, and recovery. This keeps language-model inference out of the latency-critical path. Speech-to-text and text-to-speech also run on-device alongside the game client. The agent uses a context budget of roughly 5,000 tokens and compaction to carry unfinished plans between agent loops.

Training required real gameplay because player and Ally continually shape each other’s behavior. Over 28 days at a rented gaming cafe in Korea, 1,046 participants played 38,956 sessions with Ally, averaging 14.1 minutes and 59 trajectories per session. The team first used Gemma 4 31B as a teacher, with its prompt optimized by GEPA, then generated teacher corrections for selected student rollouts. It trained an intermediate 8B teacher and distilled a deployable 2B student. The initial corpus contained 464K examples: 420K teacher rollouts and 44K synthetic and safety examples; correction sets added 313K examples.

A concrete sequence shows the mechanism. When the player is downed, Ally can observe one enemy at 2 o’clock, check that smoke and revive are available, say “One at 2 o’clock. Smoking now. Coming to get you!”, deploy smoke, move to the player, and revive. This links observation, speech, and action rather than treating them separately.

In the two-week multilingual live beta supporting English, Korean, and Chinese, a survey reached players in 141 countries. Among respondents whose play with Ally was confirmed in game records, positive responses exceeded negative responses by 25.1 percentage points when asked whether they would recommend Ally. Separately, 18.5% selected “teammate” and 31.5% selected companion framings, together 50.0% of respondents. The paper reports a single spoken exchange completed in about 1.6 seconds on-device, versus 3.4 seconds with the cloud configuration.

The approach has limits. On-device compute remains insufficient for language-model inference at game-tick rate, which is why the behavior tree exists. Ally shares memory and compute with PUBG, and the deployed configuration targets consumer GPUs with at least 8 GB of VRAM. The study is scoped to one human teammate on Sanhok, and the paper does not establish that Ally improves match outcomes or would generalize to other games. Still, the work offers a detailed, player-tested blueprint for conversational embodied agents that act alongside people in real time.