FlowWAM

Checkpoints for FlowWAM — a dual-stream (RGB + optical-flow) Wan2.2 world model paired with an IDM action expert for RoboTwin dual-arm manipulation.

• Paper: FlowWAM: Optical Flow as a Unified Action Representation for World Action Models
• Project Page: https://flow-wam.github.io/
• Code: https://github.com/YixiangChen515/FlowWAM
• Dataset: https://huggingface.co/datasets/YixiangChen/FlowWAM_RoboTwin

Files

File	Description
flowwam_robotwin.safetensors	RoboTwin dual-stream flow-action IDM checkpoint (video DiT + flow_stream + action expert).
flowwam_robotwin_action_norm_stats.npz	14-D action normalization stats (mean/std). Required for inference.
flowwam_worldarena_stage1.safetensors	FlowWAM WorldArena world-model checkpoint.

RoboTwin flow-action IDM checkpoint

flowwam_robotwin.safetensors jointly denoises an RGB stream and a head-camera optical-flow stream, and predicts a 14-D dual-arm action chunk by cross-attending to the video DiT's per-layer hidden states. It is trained on the RoboTwin 2.0 aloha-agilex demonstrations (see the dataset repo above).

Inference

Keep flowwam_robotwin.safetensors and flowwam_robotwin_action_norm_stats.npz in the same folder — the server derives the norm-stats path from the checkpoint path. The decoding config must match the training config:

Setting	Value
num_frames	33
num_video_frames	9
num_action_layers	30
action_pred_target	velocity
action_pos_mode	rope
proprio_mode	text
action_snr_shift	5.0
action_dim	14

# From the FlowWAM code repo (https://github.com/YixiangChen515/FlowWAM)
cd inference
CHECKPOINT=/path/to/flowwam_robotwin.safetensors bash start_server.sh

See the code repo's inference/README.md for the full RoboTwin evaluation setup.

Citation

@misc{flowwam,
      title={FlowWAM: Optical Flow as a Unified Action Representation for World Action Models}, 
      author={Yixiang Chen and Peiyan Li and Yuan Xu and Qisen Ma and Jiabing Yang and Kai Wang and Jianhua Yang and Dong An and He Guan and Gaoteng Liu and Jianlou Si and Jun Huang and Jing Liu and Nianfeng Liu and Yan Huang and Liang Wang},
      year={2026},
      eprint={2607.13017},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2607.13017}, 
}

License

Released under the Apache-2.0 License.