Virevium Research

Research

Virevium One

A locally run, supervised agent that plays a game the way a person does: by looking at the screen and pressing keys. It is the first Virevium project and it is in active development.

The question

Can an agent learn from its own experience to play alongside people as a useful teammate, and can each step toward that be shown with evidence someone else can check? Virevium One builds that up one capability at a time: reaching a place, handling a threat, travelling at speed.

How it works today

  • Eyes

    It reads the game window as pixels. Position and heading come from the game's own on-screen debug text; enemy health bars are found by colour and shape.

  • Body

    It sends ordinary keyboard and mouse input. It does not read the game's memory, inject code or edit game files.

  • Skills

    Local control routines for navigation, self-defence and mounted travel, each built and tested one at a time.

  • Direction

    Between runs, an AI assistant chooses goals and waypoints with the founder supervising. Nothing in the cloud acts inside the real-time loop.

  • Record

    Every decision tick is logged with its inputs, and frames are saved, so a run can be examined afterwards.

Where it runs

The proving ground is Veloren, an open-source game, in a private single-player world on one PC. Virevium is not affiliated with the Veloren project.

Demonstrated so far

Each of these is documented in the progress journal as a development note, with its method and evidence.

DateSkillWhat was measured
2026-11-10Walk to a point3 of 3 round trips on flat ground and 3 of 3 on steps, each ending within 3 units of the target
2026-11-10Follow a road298 units each way, 40 s out and 36 s back, one stall cleared by a jump
2026-11-10Defend itselfDetected an attack and defeated the attacker in two live engagements
2026-11-10Mount and dismount10 of 10 cycles confirmed by the agent's own check
2026-11-10Ride to a map destinationReached a destination about 500 units away on horseback, with one operator-assisted detour

The journal has each of these in full.

From the recorded runs

The game character standing on stone steps in daylight, with the game's debug text in the top left corner and a minimap in the top right.
The end of a walk-to-a-point run on steps. The text at the top left is the game's own debug overlay, which the agent reads for its position and heading.
Two frames of the character fighting a rat: a sword swing, then an experience award floating above the ground.
Two frames from the second rat engagement: a strike, then the experience award after the rat was defeated.
Two frames: the character riding a horse, then standing beside it with the game's Mount prompt showing.
Cycle 1 of 10 in the mount test: riding, then back on foot beside the horse, where the game shows its Mount prompt.

Frames are from the project's own recorded runs in Veloren, an open-source game.

In active development

Adaptive navigation
Recovering from unfamiliar obstacles without operator input.
Sustained engagements
Healing, retreat and tougher opponents.
Learning from experience
A standard episode format for the logs, then the first learned components trained and tested on them.
Cooperative play
Working alongside human players in approved multiplayer environments, under supervision.
Scored trials
Moving each capability from development note to a result scored against rules fixed beforehand.