RynnUnify-WAM-RoboCasa

RynnUnify-WAM-RoboCasa is the action policy model (one-step flow pose, RoboCasa ablation framework, iteration 8664) from the RynnUnify family, exported in Hugging Face safetensors format from a PyTorch distributed checkpoint (DCP).

• Architecture: Cosmos3OmniModel with action_gen (one-step flow pose action head + modality adapters)
• Contains: language model, one-step-flow-pose action head, vision/action/state adapters
• Precision: bfloat16 safetensors

Code and inference

The model definition and serving code that this checkpoint requires are published in code/ — a snapshot of the cosmos-framework fork containing the one-step flow-pose model class, action head, CFG readout, RoboCasa policy server (cosmos_framework.scripts.action_policy_server_robocasa_one_step_flow_pose) and the OpenPI evaluation client.

The released checkpoint is cfg1: the action head reads out after the generator update, so --guidance 3 is supported (generator NFE=3). Run with --num-steps 1 --shift 1 --execute-steps 16.

trust_remote_code=True is not supported: config.json records internal aliases (one_step_flow_pose_model, wan2pt2_vae_interface, ...) that are registered by the code in code/. See code/README.md for required external assets (Qwen3-VL tokenizer/processor, Wan2.2 VAE, reason-memory donor, action/state stats) and the server command.

Part of the RynnUnify collection: RynnUnify-Base | RynnUnify-Agentic-Reasoner | RynnUnify-WAM-RoboCasa

Citation

@misc{rynnunify2026,
  title={RynnUnify: Towards Unified Embodied Brain},
  author={Minghao Zhu and Bohan Hou and Kun Xiang and Zixuan Wang and Yinzhou Tang and Yifei Yang and Zhikai Wang and Tianyi Zhang and Kehan Li and Haoyu Zhao and Siteng Huang and Leigang Qu and Sicong Leng and Dongchi Huang and Mingxiu Chen and Shijian Lu and Yue Wang and Xiaodan Liang and Zhenyu Zeng and Deli Zhao and Chengju Liu and Jianfei Yang and Xin Li},
  year={2026},
}