RDT-1B

Tsinghua University (TSAIL) · 清华大学

Embodied / robotics model — runs on robot hardware, not a token API. No per-1M pricing. For API-priced LLMs see the model catalog.

At a glance

CreatorTsinghua University (TSAIL) (清华大学)
ArchitectureDiffusion Transformer (Diffusion)
EmbodimentBimanual manipulation
Parameters1.2B
LicenseMIT
Released2024-10-10

Overview

RDT-1B is a 1.2B-parameter diffusion-based foundation model for bimanual robotic manipulation, from Tsinghua's TSAIL group. Pre-trained on 1M+ multi-robot episodes, it predicts the next 64 robot actions from a language instruction plus 3-view RGB observations. Published at ICLR 2025. Both code and weights are MIT-licensed — the most permissive license among the major Chinese embodied models, which makes it a common research baseline. A newer RDT2 line is also published.

What it's used for

Bimanual manipulation research; permissive-license baseline for fine-tuning on custom robot data.

Primary sources

Other Chinese embodied models