Skip to content
RobothecaBeta
Log in

Motus

  • AI model
  • Humanoids and robot learning

Motus is a unified latent action world model that combines video generation, understanding and action experts to predict robot behavior, with checkpoints for RoboTwin 2.0 tasks.

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/Motus (opens in a new tab). Facts read at commit 88f6fe3 (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 88f6fe3 (opens in a new tab), on 16 Oct 2026.

Commercial use
Yes

What it does

  • Uses a Mixture-of-Transformers architecture with three experts: understanding, action, and video generation.
  • A UniDiffuser-style scheduler lets it switch between world model, VLA, inverse dynamics, video generation, and video-action prediction modes.
  • Learns latent actions from optical flow to extract pixel-level delta actions for large-scale action pretraining.
  • Built on Wan2.2-5B for video generation and Qwen3-VL-2B for vision-language understanding, about 8B parameters total.
  • Reports an 87.02% average success rate on RoboTwin 2.0 simulation with multi-task training across 50 tasks.
  • Provides three checkpoints covering pretraining, latent action pretraining, and RoboTwin2 fine-tuning, plus inference and training code.

Written by Robotheca from the project's README (opens in a new tab) at commit 88f6fe3.

Getting started

Difficulty
Advanced

Maintenance

Status
Active: a commit in the last 90 days
Newest commit
9 Oct 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.