RDT-1B: A Diffusion Foundation Model for Robotic Manipulation

Songming Liu, Lingxuan Wu, Bangguo Li, Hengkai Tan, Huayu Chen, Zhengyi Wang, Ke Xu, Hang Su, Jun Zhu

Tsinghua University · Shanghai AI Lab

Abstract

We introduce RDT (Robotics Diffusion Transformer), a diffusion foundation model for robotic manipulation. RDT is pre-trained on the largest multi-robot dataset to date (DROID, 1.0M+ trajectories) and then fine-tuned on a target robot. With 1.2B parameters, RDT significantly outperforms existing methods such as Diffusion Policy and OpenVLA across 5 benchmarks and 60+ tasks, achieving a 14× increase in success rate on dexterous tasks and demonstrating strong generalization to unseen objects, scenes, and instructions.

1. Introduction

Recent advances in generative modeling have unlocked new capabilities in language and vision, yet robotics still relies heavily on task-specific imitation learning. We argue that the lack of large-scale, diverse robot data — not model architecture — is the primary bottleneck. RDT tackles this by training a 1.2B-parameter transformer-based diffusion model on the DROID corpus, then transferring to downstream robots via lightweight fine-tuning (≈ 2 GPU-hours).