RDT (Robotics Diffusion Transformer)
Tsinghua Machine Learning GroupRDT is Tsinghua University's line of bimanual robot foundation models. RDT-1B is a one-billion-parameter diffusion transformer pretrained on more than a million multi-robot episodes; RDT2 adapts Qwen2.5-VL-7B into a vision-language-action model trained on over ten thousand hours of data from a handheld gripper, and deploys zero-shot on bimanual UR5e and Franka arms it was not trained on.
Openness
3 high confidence- weights
- open(RDT2 and RDT-1B checkpoints on the Hub)
- data
- described(RDT2's 10,000+ hours of UMI data described, not released)
- code
- open(training and fine-tuning scripts)
- license
- Apache-2.0(RDT2)+MIT(RDT-1B)
Weights and training code are released under Apache-2.0 and MIT. The first release also published its fine-tuning data, but the ten thousand hours of gripper recordings that RDT2 learned from are described and not published.
- https://huggingface.co/api/models/robotics-diffusion-transformer/RDT2-VQ?expand[]=siblings recorded 2026-09-26
Hub API for RDT2-VQ: four safetensors shards, ungated.
- https://huggingface.co/robotics-diffusion-transformer/rdt-1b/raw/main/README.md recorded 2026-09-26
rdt-1b card: "All the code, pre-trained model weights, and data are licensed under the MIT license"; pretrained on 46 datasets including RT-1, RH20T, DROID and BridgeData V2.
- https://huggingface.co/robotics-diffusion-transformer/RDT2-VQ/raw/main/README.md recorded 2026-09-26
RDT2-VQ card: license apache-2.0, base_model Qwen/Qwen2.5-VL-7B-Instruct.
- https://raw.githubusercontent.com/thu-ml/RDT2/main/LICENSE recorded 2026-09-26
RDT2 LICENSE: Apache License 2.0, "Copyright [2025] [TSAIL Group, Tsinghua University]".
- https://raw.githubusercontent.com/thu-ml/RDT2/main/README.md recorded 2026-09-26
RDT2 README: trained on "10,000+ hours" of UMI gripper data across "100+ different indoor scenes"; no dataset release for it.
- https://ungh.cc/repos/thu-ml/RDT2/files/main recorded 2026-09-26
RDT2 tree: rdt/train.py, vla_trainer.py, scripts/finetune_full_param.sh, finetune_lora.sh.
Adoption
1 medium confidenceHugging Face downloads summed over the RDT2 and RDT-1B checkpoints.
- https://huggingface.co/api/models?author=robotics-diffusion-transformer&sort=downloads&limit=40&expand[]=downloads&expand[]=likes&expand[]=cardData&expand[]=lastModified&expand[]=gated recorded 2026-09-26
Hub listing for robotics-diffusion-transformer: rdt-1b 546, RDT2-VQ 218, RDT2-FM 129, rdt-170m 115 downloads in 30 days.
Capability
4 low confidenceRDT learns from many robots and is built around two-armed manipulation, and RDT2 claims to run on bimanual arms it never saw in training, which few open policies attempt. Its tasks are simple ones such as picking, placing and wiping.
- https://raw.githubusercontent.com/thu-ml/RDT2/main/README.md recorded 2026-09-26
RDT2 README: "the first foundation model that can achieve zero-shot deployment on unseen embodiments for simple open-vocabulary tasks like picking, placing, shaking, wiping"; platforms "Bimanual UR5e" and "Bimanual Franka Research 3".
- https://raw.githubusercontent.com/thu-ml/RoboticsDiffusionTransformer/main/README.md recorded 2026-09-26
RDT-1B README: "1B-parameter imitation learning Diffusion Transformer pre-trained on 1M+ multi-robot episodes".
Verified 2026-09-26