Tsinghua RDT-1B

2024-10-10 · Publisher : Université Tsinghua · Domain : Robotics

Tsinghua University releases RDT-1B, a 1.2-billion-parameter diffusion transformer for two-arm manipulation, pretrained on the largest multi-robot data collection of its time.

### Background On October 10, 2024, a Tsinghua University team (Beijing) released RDT-1B (Robotics Diffusion Transformer), a foundation model dedicated to bimanual manipulation. Coordinating two arms is hard: many actions are plausible and data is scarce. ### Specifications - **Architecture**: a diffusion transformer that can represent several plausible actions for the same situation, designed for heterogeneous, high-frequency robot data. - **Size**: 1.2 billion parameters, presented as the largest diffusion-based foundation model for robot manipulation at release. - **Unified action space**: a "physically interpretable" representation that harmonizes actions across robots while preserving their physical meaning. - **Data**: pretrained on the largest multi-robot dataset collection of the time, then fine-tuned on 6,000+ bimanual episodes collected by the team. - **License**: weights released on Hugging Face under the MIT license. ### Why it matters RDT-1B shows that diffusion, already dominant in image generation, adapts to coordinated two-arm motion. It also makes China a visible contributor of open robot models. ### Key takeaway RDT-1B became a bimanual baseline: AgiBot compared its GO-1 model against it in 2025.

Tags : Diffusion, Bimanuel, Open Source, Modèle de fondation