Project Page | Paper | Twitter | YouTube

Ge Yan, Jiyue Zhu*, Yuquan Deng*, Shiqi Yang, Ri-Zhao Qiu, Xuxin Cheng, Marius Memmel, Ranjay Krishna†, Ankit Goyal†, Xiaolong Wang†, Dieter Fox†

Conference on Robot Learning (CoRL) 2025

ManiFlow is a visuomotor imitation learning policy for general robot manipulation that generates precise, high-dimensional, and dexterous actions from visual, language, and proprioceptive inputs. It features a general architecture supporting both 2D and 3D inputs, combines flow matching with consistency training for 1-2 step inference, and demonstrates strong scaling behavior and generalization across diverse robots and environments.

• Superior Performance across 60+ simulation tasks on 4 benchmarks, 98.3% improvement on 8 real-world tasks with strong generalization to novel objects, backgrounds, and cluttered scenes.

• Diverse Robot Embodiments: Single policy works across diverse platforms with increasing dexterity - humanoid robots (Unitree H1), bimanual systems (dual xArm), and single-arm setups (Franka Panda)

• Universal Policy Architecture: Unified framework handles both 2D RGB images and 3D point clouds.

• Strong Scaling Capability: Shows superior data efficiency and scaling performance, achieving 99.7% success on a single task with 500 demos while outperforming large-scale pre-trained π0 model.

• Few-Step Inference: Generates precise dexterous actions in just 1-2 inference steps vs. 10+ steps for diffusion baselines through consistency flow training in a single run without pre-trained teacher models.

• General Applicability: ManiFlow's consistency flow training objective can be seamlessly integrated into existing VLA model by replacing the diffusion loss, enabling faster inference and improved performance.

• 🚀2025-11-12, Support RoboTwin2.0! Please check RoboTwin2.0 branch for more details.

• ❗2025-09-03, Released ManiFlow, a general robot manipulation policy via consistency flow training. Check out the Project Page and Paper.

• 🚀 More efficient and capable 3D encoders for large-scale scenes with dense point clouds input
• 🤖 Real-Time Execution with RTC and varied temporal ensembling

• Installation
• Data Preparation
• Training & Evaluation
• Practice Tips
• Real-World Deployment
• Q&A
• Adding Your Own Method
• License
• Acknowledgements
• Citation
• Contact

Please follow the detailed instructions in INSTALL.md to set up the environment and install dependencies.

The scripts for generating data are provided in the scripts/ directory. Generated data will be stored in the ManiFlow_Policy/ManiFlow/data/ directory by default as .zarr format files for efficient data loading.

RoboTwin provides a comprehensive suite of bimanual dexterous manipulation tasks featuring visually realistic environments and diverse scene configurations. The benchmark includes complex scenarios with varying object properties, lighting conditions, and environmental clutter to test policy robustness and generalization.

🔹 Current Support - RoboTwin 1.0: Generate demonstrations for the example diverse_bottles_pick task using the provided scripts:

bash scripts/gen_demonstrations_RoboTwin1.0.sh diverse_bottles_pick 0

⬇️ Download pre-generated data: For convenience, pre-generated RoboTwin 1.0 datasets are available on Google Drive. Download and extract the .zarr files directly to your data directory.

🔥 RoboTwin 2.0: Please check RoboTwin2.0 branch for more details, with enhanced task diversity, comprehensive domain randomization, and more challenging manipulation scenarios.

These three simulation environments provide diverse manipulation tasks for training and evaluation. Each environment offers different types of challenges:

• Adroit: Hand manipulation tasks with dexterous control
• DexArt: Complex dexterous manipulation with articulated objects
• MetaWorld: Multi-task manipulation benchmark with 50+ tasks

Generate demonstrations for these environments using the corresponding scripts in the scripts/ directory.

Please refer to the individual scripts for specific task names and available parameters. Each script allows customization of demonstration count, environment settings, and data collection parameters.

ManiFlow supports both 2D RGB and 3D point cloud policy training across different benchmarks. Use the provided training scripts in the scripts/ directory to train policies:

Train on RoboTwin Environment:

1. Train a 3D ManiFlow Policy for a single task with seed 0 on GPU 1 in the robotwin environment:

   bash scripts/train_eval_robotwin.sh maniflow_pointcloud_policy_robotwin pick_apple_messy_pointcloud 0901 0 1

2. Train a 2D ManiFlow Policy using the same script:

   bash scripts/train_eval_robotwin.sh maniflow_image_timm_policy_robotwin pick_apple_messy_image 0901 0 1

Train on Adroit and DexArt Environments:

For Adroit and DexArt environments, use the same training scripts with different environment and task names. For example:

1. Adroit:

   bash scripts/train_eval_dex.sh maniflow_pointcloud_policy_dex adroit_door_pointcloud 0901 0 1

2. DexArt:

   bash scripts/train_eval_dex.sh maniflow_pointcloud_policy_dex dexart_laptop_pointcloud 0901 0 1

Train on MetaWorld Environment:

Train a language-conditioned multi-task ManiFlow policy on MetaWorld with multiple GPUs:

bash scripts/train_eval_metaworld.sh maniflow_pointcloud_policy_metaworld metaworld_multitask_mp debug 0 1_2_3

Multi-GPU Configuration:

• Set GPU IDs as a list format (e.g., 0_1_2) in the training scripts to utilize multiple GPUs.
• Evaluation runs automatically on multiple GPUs to accelerate the evaluation process.
• Use eval_env_processes parameter in scripts to control the number of parallel evaluation environments across all GPUs.

Automatic Evaluation: Set eval=True in RoboTwin and MetaWorld training scripts to automatically evaluate the saved trained model on the specified task after training completion.

Online Evaluation: Adroit and DexArt environments use online evaluation during training with results logged to wandb.

Customization: Check individual scripts for detailed parameter descriptions and customization options.

• 🔍 Dense representation matters. Avoid pooling operation in the point cloud encoder to preserve fine-grained geometric details. Use 128-256 points for well-calibrated cropped scenes, and 2048-4096 points for cluttered environments.
• 🌈 Apply color jitter (brightness, contrast, saturation) with 0.2 probability during training to prevent overfitting to specific lighting conditions and improve real-world robustness.
• 👁️ For active sensing cameras with dynamic viewpoints, collect demonstrations with head movements to enable coordinated head-arm control for humanoid platforms and better generalization to novel viewpoints.
• 🔧 Use temporal ensembling for real-world deployment to ensure safety, smoother transitions and account for execution delays.
• ⏱️ Use longer action horizons (16+ steps) for complex long-horizon tasks to capture full temporal dynamics.
• 📈 Data diversity is crucial for generalization. Collect demonstrations with varying object properties, backgrounds, and clutter to improve robustness.

ManiFlow has been successfully deployed across three distinct real-world robot platforms, demonstrating superior performance in challenging manipulation tasks with increasing dexterity requirements.

• Hardware: Full-sized humanoid with 28-DoF (7-DoF arms + 6-DoF anthropomorphic Inspire hands + 2-DoF active head). Use upper-body only for manipulation tasks.
• Perception: Gimbal-mounted ZED stereo camera for active egocentric sensing
• Teleoperation: Apple Vision Pro - refer to OpenTeleVision for setup
• Key Features:
  • Coordinated head-arm movements for dynamic viewpoint handling
  • Multi-finger coordination with 6-DoF per anthropomorphic hand
  • Active sensing capabilities with moving camera perspectives

• Hardware: Two UFACTORY xArm 7 robotic arms with PSYONIC Ability Hands (26-DoF total)
• Perception: Intel RealSense LiDAR L515 camera with front-facing static view
• Teleoperation: Apple Vision Pro - refer to Bunny-VisionPro for setup
• Key Features:
  • Precise bimanual coordination and synchronization
  • Dexterous manipulation with 6-DoF hands
  • Consistent third-person visual perspective

• Hardware: 7-DoF Franka Emika Panda with Robotiq parallel gripper
• Perception: Intel RealSense D455 RGB-D camera (statically mounted)
• Teleoperation: Oculus VR for data collection
• Key Features:
  • Industrial-grade precision and reliability
  • External visual observations for consistent viewpoint

Q: What is the limitation of 3D perception in ManiFlow?

A: ManiFlow currently leverages dense point cloud representation, which can be computationally expensive for larger scale scenes requiring more points. Future work could explore more efficient 3D encoder to compress the point cloud while preserving geometric details.

Q: How much demonstration data does ManiFlow need?

A: For simple tasks, ManiFlow can achieve reasonable performance with as few as 20-50 demos. For complex long-horizon tasks, 100-200 demos are recommended for optimal performance. ManiFlow shows strong scaling behavior, so more data generally leads to better results.

To integrate your custom manipulation policy:

1. Implement your policy inheriting from BasePolicy in maniflow/policy/, e.g., maniflow/policy/maniflow_pointcloud_policy.py.
2. Register policy config in config/policy/, e.g., maniflow/config/maniflow_pointcloud_policy_robotwin.yaml.
3. Train using existing scripts: bash scripts/train_eval_*.sh your_method task_name

This project is licensed under the MIT License - see the LICENSE file for details.

We thank the authors of the following open-source projects, whose code were invaluable in developing ManiFlow: Diffusion Policy, UMI, 3D Diffusion Policy, Improved 3D Diffusion Policy, RDT, RoboTwin, consistency models, shortcut models

If you find this repository useful for your research, please consider citing the following paper:

@inproceedings{yan2025maniflow,
  title={{ManiFlow}: A General Robot Manipulation Policy via Consistency Flow Training},
  author={Yan, Ge and Zhu, Jiyue and Deng, Yuquan and Yang, Shiqi and Qiu, Ri-Zhao and Cheng, Xuxin and Memmel, Marius and Krishna, Ranjay and Goyal, Ankit and Wang, Xiaolong and Fox, Dieter},
  booktitle={Conference on Robot Learning (CoRL)},
  year={2025}
}

For questions, issues, or collaboration inquiries:

• 🐛 Open an issue in this repository
• 💬 Reach out directly to Ge Yan