survev
survev is an open-source clone of the game surviv.io, supporting multiplayer battle royale gameplay. It is frequently used as a research environment for multi-agent reinforcement learning, providing a controllable testbed for studying competition, cooperation, and emergent behavior among learning agents.
Timeline (last 90 days)
On Reddit, a developer shared an experiment using MAPPO to train a 4-person squad to play survev (an open-source surviv.io clone), and released a 40-second gameplay clip.
survev is an open-source surviv.io clone in which every character in a match is controlled by an AI bot.
In the experiment, the four squad members share the same actor network, with agent id used as input to distinguish identities.
The actor's input is strictly identical to that of the real game client: about 1.4k-dimensional features, two 24x24 local grids, and a minimap, processed by a small CNN, with set encoding for up to 12 visible enemies, then fed into a 256-dimensional GRU for temporal memory.
The actor's input is consistent with that of the real client, so the same network can be packaged as a browser extension and run directly, without needing a server backdoor.
Privileged state (global enemy positions, full health overview, etc.) is fed only to the centralized critic, to reduce variance during training.
The action space is divided into three layers: first select intent (looting, exploring, finding cover, peeking, retreating, regrouping, pushing, entering the zone, etc.), then select target, and finally select the specific action (gun, melee, grenade, smoke, healing).
The squad shares a single team reward, and the reward terms include damage dealt and taken, knockdowns and kills, final ranking, plus team cooperation terms such as focus fire, crossfire, and rescue.
The author did not disclose the code and all details, on the grounds that once disclosed, this kind of bot could be used for cheating on public servers.
All Facts (9)
The author did not disclose the code and all details, on the grounds that once disclosed, this kind of bot could be used for cheating on public servers.
50%Unverifiedsurvev is an open-source surviv.io clone in which every character in a match is controlled by an AI bot.
50%UnverifiedIn the experiment, the four squad members share the same actor network, with agent id used as input to distinguish identities.
50%UnverifiedThe actor's input is strictly identical to that of the real game client: about 1.4k-dimensional features, two 24x24 local grids, and a minimap, processed by a small CNN, with set encoding for up to 12 visible enemies, then fed into a 256-dimensional GRU for temporal memory.
50%UnverifiedThe actor's input is consistent with that of the real client, so the same network can be packaged as a browser extension and run directly, without needing a server backdoor.
50%UnverifiedPrivileged state (global enemy positions, full health overview, etc.) is fed only to the centralized critic, to reduce variance during training.
50%UnverifiedThe action space is divided into three layers: first select intent (looting, exploring, finding cover, peeking, retreating, regrouping, pushing, entering the zone, etc.), then select target, and finally select the specific action (gun, melee, grenade, smoke, healing).
50%UnverifiedThe squad shares a single team reward, and the reward terms include damage dealt and taken, knockdowns and kills, final ranking, plus team cooperation terms such as focus fire, crossfire, and rescue.
50%UnverifiedOn Reddit, a developer shared an experiment using MAPPO to train a 4-person squad to play survev (an open-source surviv.io clone), and released a 40-second gameplay clip.
50%