AI in Dota 2: Dota2-Vis studies when the enemy can see your hero

What the AI actually measures

The researchers view each match from both teams’ perspectives. A hero’s position on the full map and the information available to the enemy are different things. Fog of war, trees and vision affect what can be seen. The project studies a hero’s presence on the enemy minimap; it does not read the player’s intentions.

Showing on a lane is not necessarily a mistake

A carry may deliberately push a lane while the team takes control elsewhere. A support may show to clear a dangerous wave. These are our gameplay examples: a visibility graph alone does not prove bad positioning or a good move. You need the team’s objective and the match situation.

How to use the idea in your own replay

Pick one death before an important fight. In the replay, switch to the enemy’s vision and find the last time your hero appeared: where did they see you, and what route could they have inferred afterwards? Do not replace the information available to a player with the observer’s all-seeing view.

Then choose one adjustment: a different route, waiting for information, or showing on a lane at a different time. This is an OOPS DOTA editorial exercise; we are not claiming to have tested the software or proven an MMR increase.

FAQ: can you download a finished coach yet?

We cannot confirm a full public launch. The paper describes the dataset as available, but the repository README still promises code and access instructions later. A link to the research should therefore not be mistaken for a ready-to-use app for your matches.

FAQ: is this a continuation of OpenAI Five?

This is a different project. In 2019, OpenAI Five beat OG 2:0 in a historical experiment with restricted rules. Five controlled heroes; Dota2-Vis studies visibility in recordings. The result of that old experiment does not describe what a modern chatbot can do.