RDT-1B

Overview

RDT-1B is a robot-pretraining research project associated in the source with 刘松明 and his collaborators before the founding of Liber AI.

Current Profile

The interview presents RDT-1B as evidence that robot tasks share transferable capability: a pretrained model followed by a small amount of task-specific post-training reportedly outperformed training only on the post-training data, including on unseen tasks. The project used a diffusion approach suited to continuous actions and published detailed training and failure lessons, which helped create a large technical community around the work.

Its strategic importance in this source is not that it solved general robot manipulation. It clarified the next bottleneck: producing diverse, precise, high-quality robot-action data cheaply enough for pretraining to continue scaling.

Key Characteristics

• Large-scale robot pretraining intended to transfer capability across tasks.
• Diffusion-based modeling chosen for continuous action data.
• Evaluation emphasis on the value of pretraining before small task-specific post-training.
• Detailed open technical documentation and failure reporting.
• Community formation that connected researchers, industrial practitioners, and later startup collaborators.

Evidence

• Transfer and post-training: 187: 对话LiberAI刘松铭：Astra来袭，00后物理AI创始人看到的胜率 reports that pretrained models with limited post-training data beat post-training-only baselines on unseen tasks.
• Method and disclosure: 187: 对话LiberAI刘松铭：Astra来袭，00后物理AI创始人看到的胜率 attributes the diffusion choice and detailed open training lessons to the project.
• Research-to-company bridge: 187: 对话LiberAI刘松铭：Astra来袭，00后物理AI创始人看到的胜率 describes the project’s discussion groups and industry feedback as inputs to Liu’s later founder network and data strategy.

Qualifications

The supplied source does not provide the paper, repository, benchmark protocol, contributor list, training corpus, or numerical results needed to audit RDT-1B directly. All technical and community-impact claims on this page are bounded to the interview and should be checked against primary project materials if later ingested.

What Changed

• Created the project page from its role in Liu’s research and company formation.
• Separated the reported transfer result from the unresolved robot-data scaling bottleneck.

Relationships

• 刘松明 / Liu Songming - researcher and later founder associated with the project in the source.
• Liber AI - later company whose pretraining and data strategy draws on this research lineage.
• Robot Data Scale Up - bottleneck the project reportedly made more visible.
• Vision Language Action Models - adjacent robot-policy family discussed in the same research period.
• World Action Models - later action-conditioned route pursued by Liber AI.