RDT-1B

Um Tsinghua University (TSAIL / IIIS)

Stofnað 1911 · Beijing, China

Robotics Diffusion Transformer (RDT) is a generalist bimanual manipulation policy developed at Tsinghua University's TSAIL / Institute for Interdisciplinary Information Sciences (IIIS), home of Jun Zhu's diffusion-modelling group. RDT-1B, introduced in October 2024, is one of the first publicly released billion-scale diffusion-based Vision-Language-Action models, specifically designed for two-arm robots such as Aloha, Mobile Aloha and a custom bimanual platform used by the authors. The project is positioned as a Chinese academic counterpart to π-0 and OpenVLA, with open weights released on Hugging Face under a permissive licence and the explicit aim of enabling fully reproducible bimanual VLA research.

Arkitektúr

Diffusion Transformer Vision-Language-Action policy for bimanual manipulation

RDT-1B is a 1-billion-parameter Diffusion Transformer (DiT) trained as a Vision-Language-Action policy. Inputs are multi-view RGB observations (left + right + overhead), proprioception for both arms and any gripper / mobile-base degrees of freedom, plus a natural-language instruction encoded by a text encoder. The conditioning tokens are fed through a transformer trunk, while a diffusion head denoises continuous action chunks for both arms in a unified action space, allowing dual-arm coordinated motion. Pretraining is done in two stages: a large multi-robot pretraining phase on >1M episodes drawn from public datasets including Open-X-Embodiment and curated bimanual corpora, followed by fine-tuning on the authors' own 6,000-episode bimanual dataset spanning ~300 tasks. RDT-1B reports strong results on dexterous bimanual tasks such as folding T-shirts, pouring, and tool use.

Breytur

1B

Geta

• 1B-parameter Diffusion Transformer VLA
• Designed for bimanual manipulation (Aloha-class robots)
• Trained on >1M cross-embodiment episodes + 6k bimanual demos
• Continuous action chunks for both arms in a unified space
• Diffusion head produces smooth coordinated motion
• Open weights on Hugging Face (permissive licence)
• Strong results on folding, pouring and tool use
• Reproducible training and evaluation code
• Best for: bimanual manipulation research, two-arm fine-tuning.

Þjálfun og leyfi

Pretraining on >1 million robot episodes from Open-X-Embodiment and other public datasets, followed by fine-tuning on a curated bimanual dataset of ~6,000 episodes covering ~300 tasks collected with Aloha-class hardware.

Leyfi: Open weights released on Hugging Face under a permissive (CC-BY-NC-style) licence; primarily intended for research use.

Öryggispróf: Academic research artifact; no formal red-teaming or RSP. Safety is the responsibility of downstream deployers (Aloha hardware already includes torque limits and e-stop).