We introduce PUBG Ally, an embodied agent that can team up with players in PUBG: BATTLEGROUNDS. It perceives the constantly changing game world, responds under strict latency constraints, and synchronizes its speech with in‑game actions.
Ally is built on a language‑model agent that uses a controlled interface to read game state, interpret player speech, keep conversational context, decide what to say, and emit high‑level action choices. These choices steer a faster control layer that handles movement, combat and recovery, thus merging tool use with real‑time game control.
To teach the agent to cooperate with humans, we collected data from almost 39 k real matches. Each session records gameplay video, player voice, agent decisions, tool usage, concrete actions, and player feedback, which we use for iterative training.
For evaluation we combine player feedback and preference comparisons, uncovering gaps between offline metrics and actual player preferences. The evaluation criteria are refined iteratively to better reflect user experience.
Deploying Ally in live service demands low‑latency on‑device inference and safe player‑facing communication. We address these through model compression, context compaction, targeted safety training, runtime guardrails, and memory redaction.
During live service we surveyed players in 141 countries. Among those whose matches with Ally were verified, the net positive recommendation rate was 25.1 percentage points, and many described Ally not just as a tool but as a teammate or companion.
Review