GameWAM Models
GameWAM models for closed-loop game control: world action models built on Wan2.2-TI2V-5B that jointly model video prediction and native actions, and act online with block-cycle control (predict 16 steps, commit 8).
- Paper: arXiv:2608.26200
- Code: yunncheng/GameWAM
- Training data: gamewam-minecraft, gamewam-vizdoom
Repository contents
| Directory | Game | Task config | Checkpoint |
|---|---|---|---|
minecraft/ | Minecraft | mc_uncond_ar_224_4e-5 | gamewam-minecraft.pt (12.2 GB) |
vizdoom/ | ViZDoom | vizdoom_uncond_ar_224_2e-4 | gamewam-vizdoom.pt (12.2 GB) |
Each subdirectory holds one checkpoint and its paired dataset_stats.json.
Always load a checkpoint together with the dataset_stats.json from the same
subdirectory.
Minecraft model
| Base | Wan2.2-TI2V-5B (video expert, 5.0B params) + ActionDiT action expert (1.0B params) |
| Task config | mc_uncond_ar_224_4e-5 |
| Training | LR 4e-5, 21,900 steps, BF16, DeepSpeed ZeRO-2, 8xH200 |
| Proprio input | 15-D canonical state + 6-D field validity |
| Action space | 22-D keyboard/mouse, route-aware normalization, GUI head enabled |
| Control | prediction_horizon=16, execution_horizon=8 (block-cycle) |
| Precision | bfloat16 |
hf download yunncheng/gamewam-models --include "minecraft/*" --local-dir gamewam-models
bash scripts/eval_mc_multi.sh --mode all --episodes 5 \
--ckpt gamewam-models/minecraft/gamewam-minecraft.pt \
--dataset_stats gamewam-models/minecraft/dataset_stats.json \
--output evaluate_results/minecraft/gamewam-release
The default mc_uncond_ar_224_4e-5 task configuration matches this checkpoint.
ViZDoom model
| Base | Wan2.2-TI2V-5B (video expert, 5.0B params) + ActionDiT action expert (1.0B params) |
| Task config | vizdoom_uncond_ar_224_2e-4 |
| Training | LR 2e-4, 28,000 steps, bfloat16 |
| Proprio input | 4-D game state (POSITION_X, POSITION_Y, ANGLE, HEALTH) |
| Action space | 9-D (shoot, movement, strafe, speed, turn; turn_delta continuous) |
| Control | prediction_horizon=16, execution_horizon=8 (block-cycle) |
| Precision | bfloat16 |
hf download yunncheng/gamewam-models --include "vizdoom/*" --local-dir gamewam-models
bash scripts/eval_vizdoom_multi.sh --mode all --episodes 50 \
--ckpt gamewam-models/vizdoom/gamewam-vizdoom.pt \
--dataset_stats gamewam-models/vizdoom/dataset_stats.json \
--output evaluate_results/vizdoom/gamewam-release
The default vizdoom_uncond_ar_224_2e-4 task configuration matches this
checkpoint.
See the repository README for installation, data preparation, and the full evaluation protocol.
Citation
@misc{guo2026gamewamworldactionmodel,
title={GameWAM: A World Action Model for Video Games},
author={Yuncheng Guo and Zhanqiu Zhang and Yiwen Guo and Weijia Li},
year={2026},
eprint={2608.26200},
archivePrefix={arXiv},
primaryClass={cs.AI},
url={https://arxiv.org/abs/2608.26200},
}