DuoFlow

Contributor: XJTU IAIR & WisdomStride
Contact: zhengyj@stu.xjtu.edu.cn

DuoFlow is a language-conditioned policy for bimanual manipulation. This repository releases its learned adapter weights for simulation research with the Aloha-AgileX joint-action interface.

Release contents

File	Purpose
duoflow_adapter.safetensors	Learned FP32 adapter tensors: 4,046,848 parameters, rank 8, alpha 16.
adapter_config.json	Adapter format, compatible base-model revision and inference configuration.
ASSET_SOURCES.json	Pinned sources for the base model, tokenizer assets and runtime components.
LICENSE	Apache-2.0 terms for this adapter release.

The weight file contains the adapter only. Base-model weights, optimizer state, private training implementation and experiment logs are not included. The adapter uses the DuoFlow custom tensor format.

Download

from huggingface_hub import hf_hub_download

adapter_path = hf_hub_download(
    repo_id="Sva1in/DuoFlow",
    filename="duoflow_adapter.safetensors",
)

For a reproducible deployment, set revision to the commit selected from this repository's history.

Loading and compatibility

Use the DuoFlow custom inference adapter with these weights. This file is not a standalone Transformers model or a standard PEFT adapter checkpoint; adapter_config.json describes the custom format and does not imply compatibility with PeftModel.from_pretrained.

The required base model is GigaBrain-0.7-3.5B-RoboTwin2.0-Clean, pinned to revision 3de7c7e60e5779d016c8f148eed19c4b99cd2a95. Obtain the corresponding model and tokenizer files using the exact sources and revisions in ASSET_SOURCES.json.

The deployment accepts three RGB views, a language instruction and 14-dimensional joint/gripper state. It predicts 50 actions with 10 inference steps and executes up to 30 actions before refreshing observations.

The inference code and technical report are not yet publicly available. This release has not received official RoboTwin leaderboard certification.

Attribution and license

Adapter contributions: Copyright 2026 XJTU IAIR & WisdomStride. The adapter is released under Apache-2.0. Base-model, tokenizer and external runtime components retain their original licenses and attribution; this release does not relicense them. Component sources and versions are recorded in ASSET_SOURCES.json.

中文说明

本仓库提供 DuoFlow 的适配器权重，需配合固定版本的基础模型与 DuoFlow 推理代码使用，不是可独立加载的完整基础模型。本版本包含权重、配置及使用说明；训练实现未公开发布。