RDT-1B

RDT-1B is a 1-billion-parameter diffusion transformer for robot imitation learning, pre-trained on over 1 million multi-robot episodes — among the largest such datasets at release. Given a language instruction and up to three RGB camera views, it predicts the next 64 robot actions and works across single-arm to dual-arm mobile manipulators. The team fine-tuned it on 6,000+ self-collected bimanual episodes and deployed it on ALOHA dual-arm robots with state-of-the-art dexterity and few-shot results.

Limitations: research-grade robotics model requiring real robot hardware or simulation for deployment; training and inference code is PyTorch with DeepSpeed, not a turnkey product.