hqfang/sam2act-data

SAM2Act

SAM2Act is a multi-view robotics transformer policy for robotic manipulation. Built on RVT-2, it combines multi-resolution upsampling with visual embeddings from the SAM2 foundation model to improve 3D action prediction, multitask learning, and generalization. SAM2Act+ extends this policy with a memory bank, memory encoder, and memory attention so the agent can condition on prior observations and actions for spatial memory-dependent tasks.

For full project details, code, training instructions, and videos, see the SAM2Act website and GitHub repository.

Replay Buffers

This dataset repository stores pre-generated replay buffers for training SAM2Act and SAM2Act+. The buffers are serialized YARR replay buffers generated from the RLBench and MemoryBench demonstrations, so they can be loaded directly during training instead of being rebuilt on the fly.

The repository is organized as follows:

replay_temporal/replay_train/          # RLBench 18-task replay buffers
replay_temporal_memory/replay_train/   # MemoryBench replay buffers

Each task is provided as a .tar.xz archive. After downloading, extract each archive with tar -xf <task_name>.tar.xz and place the extracted task folders under the matching local directory in the SAM2Act codebase:

• RLBench: sam2act/sam2act/replay_temporal/replay_train
• MemoryBench: sam2act/sam2act/replay_temporal_memory/replay_train

These replay buffers are intended for training from scratch. They are not required for evaluating the pretrained models in hqfang/sam2act-models.

Bibtex

If you use these replay buffers, please cite the SAM2Act paper:

@misc{fang2025sam2act,
      title={SAM2Act: Integrating Visual Foundation Model with A Memory Architecture for Robotic Manipulation},
      author={Haoquan Fang and Markus Grotz and Wilbert Pumacay and Yi Ru Wang and Dieter Fox and Ranjay Krishna and Jiafei Duan},
      year={2025},
      eprint={2501.18564},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2501.18564},
}