Skip to content
RobothecaBeta
Log in

RDT-1B and RDT2

  • AI model
  • Humanoids and robot learning

RDT2 is a family of open robot foundation models that map camera images and a language instruction to robot arm actions, aiming at zero-shot use on unseen robot setups.

Made by TSAIL group, Tsinghua University, China (checked on github.com/thu-ml (opens in a new tab)).

Source on GitHub: github.com/thu-ml/RDT2 (opens in a new tab). Facts read at commit 0797b4c (opens in a new tab).

Licenses

Code
Apache-2.0 (opens in a new tab)

Read by Robotheca from LICENSE (opens in a new tab) (Apache-2.0) at commit 0797b4c (opens in a new tab), on 16 Oct 2026.

Commercial use
Yes

What it does

  • Sequel to RDT-1B, targeting simple open-vocabulary tasks such as picking, placing, shaking and wiping.
  • RDT2-VQ is a vision-language-action model built on Qwen2.5-VL-7B-Instruct, using Residual VQ as its action tokenizer.
  • Trained on over 10,000 hours of human manipulation videos collected with a redesigned UMI gripper in 100+ indoor scenes.
  • Provides checkpoints for inference and scripts for fine-tuning on your own data as WebDataset shards, with full and LoRA options.

Written by Robotheca from the project's README (opens in a new tab) at commit 0797b4c.

Getting started

Difficulty
Advanced
Hardware it needs, as the project states it
NVIDIA GPU: inference needs over 32 GB RAM and about 16 GB VRAM (e.g. RTX 4090); LoRA fine-tuning of RDT2-VQ needs over 32 GB VRAM (e.g. A100 40GB); full fine-tuning of RDT2-VQ needs over 80 GB VRAM (A100 80GB, H100 or B200); RDT2-FM expert fine-tuning needs about 16 GB VRAM. Tested on Ubuntu 24.04

Maintenance

Status
Maintained: a commit in the last 12 months
Newest commit
7 Feb 2026 on main
Latest release
None published on GitHub

Checked on GitHub on 16 Oct 2026.

Links

Robots on Robotheca that use it

No robot on Robotheca is linked to it yet. Links come from a robot's robot.yaml, from its dependency files read as text, or from Robotheca's editors.

Robotheca lists AI models as a reference: nothing is hosted or run here, and every fact links to where it was read.