delaurentis/craftax-agents

Craftax agents

Four agents for Craftax, as ONNX files. Watch them play, or take over from them, at playcraftax.com.

Agent	Network	Score / 226	Game completed	From
puffer162/	4 MinGRU layers of 1,024 (14M parameters)	162	72%	PufferLib 5.0
ppo-rnn/	one GRU of 512 (6.9M)	35	15.3%	reproduction of the PPO-RNN baseline from the Craftax paper
trxl/	gated Transformer-XL, 128-step memory (4.7M)	42	19%	Gautier Hamon, transformerXL_PPO_JAX
priml-boss/	4 MinGRU layers of 1,024 + a convolutional board encoder (15.5M)	194	86%	reproduction of rekursiv.ai's priml recipe, with its boss-fight fine-tune

Each folder has:

• <agent>.onnx: 32-bit weights
• <agent>-16bit.onnx and <agent>-8bit.onnx: smaller downloads, still computed in float32
• <agent>.safetensors: the weights for PyTorch or JAX (in priml-boss/, priml's own bf16 weights, with its tensor names so its MinGRUPolicy loads them)
• LICENSE
• in ppo-rnn/ only: params.msgpack and config.json, the original flax params
• in priml-boss/ only: metrics.json, priml's own evaluation

priml-boss beats the final boss, the necromancer, in 43% of games. Its score comes from priml's own evaluation (10,068 games on the 8,192 worlds it trained on), so expect a little less on fresh worlds.

Inputs and outputs

All four take the same input and give the same output, one game step per call:

• In: obs, 843 floats: the 9 × 11 tiles around the player (8 numbers each) plus 51 for health, inventory and the rest.
• Out: logits, a score for each of Craftax's 43 actions (rule out the ones the game doesn't allow, then take the highest), and value, the agent's guess at how well it's doing.

Each also keeps a memory from step to step. Start it at zeros and feed each step's _out back in as the next step's input:

Agent	Memory
puffer162	h [4, 1, 1024]
ppo-rnn	h [1, 1, 512]
trxl	kv [4, 1, 128, 256] and fill [1]
priml-boss	h [4, 1, 1024]

priml-boss also takes prev_action [1]: the action played on the previous step, 43 at the start of a game.

For the full format, or to convert your agent so it plays on playcraftax.com, see craftax-agents on GitHub.

All four agents are released under the MIT license, so you can use, change and share them freely. The copyright stays with each agent's authors (PufferAI for puffer162, TextJam, Inc. for ppo-rnn and priml-boss, Gautier Hamon for trxl), and the LICENSE file in each folder says so.