Motus is a unified latent action world model that leverages existing pretrained models and rich, sharable motion information. Motus introduces a Mixture-of-Transformers (MoT) architecture to integrate three experts (understanding, action, and video generation) and adopts a UniDiffuser-style scheduler to enable flexible switching between different modeling modes (World Models, Vision-Language-Action Models, Inverse Dynamics Models, Video Generation Models, and Video-Action Joint Prediction Models). Motus further leverages optical flow to learn latent actions and adopts a three-phase training pipeline and six-layer data pyramid, thereby extracting pixel-level "delta action" and enabling large-scale action pretraining.

This checkpoint is fine-tuned on RoboTwin 2.0 benchmark (50+ manipulation tasks).

Homepage | GitHub | arXiv | Feishu | WeChat

Highlights
Model Details
Performance
Hardware & Software Requirements
Quickstart (Inference)
Citation

Highlights

87.02% average success rate on RoboTwin 2.0 (+15% over X-VLA, +45% over π₀.₅)
50+ Manipulation Tasks: Trained on diverse bimanual manipulation scenarios
Multi-Task Capable: Single model handles all 50+ tasks
Ready for Deployment: Direct inference or further fine-tuning

Model Details

Architecture

Component	Base Model	Parameters
VGM (Video Generation Model)	WAN 2.2	~5.00B
VLM (Vision-Language Model)	Qwen3-VL-2B	~2.13B
Action Expert	-	~641.5M
Understanding Expert	-	~253.5M
Total	-	~8B

Training Details

Base Checkpoint: motus-robotics/Motus (Stage 2 pretrained)
Fine-Tuning Data: RoboTwin 2.0 (2,500 clean + 25,000 randomized demonstrations)
Training Steps: 40k steps

Action Representation

Control frequency: 30Hz (default)
Action chunk size: 48 steps (default)
Action dimension: 14 (bimanual: 7 per arm)

Performance

RoboTwin 2.0 Benchmark (50+ Tasks)

Method	Clean	Randomized
π₀.₅	42.98%	43.84%
X-VLA	72.80%	72.84%
Motus (Ours)	88.66%	87.02%

Key improvements:

+15% over X-VLA
+45% over π₀.₅

Hardware & Software Requirements

Mode	VRAM	Recommended GPU
Inference (with pre-encoded T5)	~ 24 GB	RTX 5090
Inference (without pre-encoded T5)	~ 41 GB	A100 (40GB) / A100 (80GB) / H100 / B200

Quickstart (Inference)

RoboTwin 2.0 Simulation

bash

cd inference/robotwin/Motus

# Single task evaluation
bash eval.sh place_dual_shoes

# Multi-task batch evaluation
bash auto_eval.sh

Offline Inference (No Environment)

bash

python inference/real_world/Motus/inference_example.py \
  --model_config inference/real_world/Motus/utils/robotwin.yml \
  --ckpt_dir ./pretrained_models/Motus_robotwin2 \
  --wan_path /path/to/pretrained_models \
  --image /path/to/input_frame.png \
  --instruction "pick up the cube and place it on the right" \
  --use_t5 \
  --output result.png

Python API

python

import torch
import yaml
from models.motus import Motus, MotusConfig

# Load config
with open("configs/robotwin.yaml", "r") as f:
    config = yaml.safe_load(f)

# Initialize model
model_config = MotusConfig(
    wan_checkpoint_path=config['model']['wan']['checkpoint_path'],
    vae_path=config['model']['wan']['vae_path'],
    wan_config_path=config['model']['wan']['config_path'],
    vlm_checkpoint_path=config['model']['vlm']['checkpoint_path'],
    action_dim=14,
    load_pretrained_backbones=False,
)

model = Motus(model_config).to("cuda").eval()
model.load_checkpoint("./pretrained_models/Motus_robotwin2", strict=False)

# Inference
with torch.no_grad():
    predicted_frames, predicted_actions = model.inference_step(
        first_frame=frame_tensor,
        state=state_tensor,
        num_inference_steps=20,
        language_embeddings=t5_embeddings,
        vlm_inputs=[vlm_inputs],
    )

# Action chunk: [1, 48, 14]
actions = predicted_actions.squeeze(0).cpu().numpy()

Citation

bibtex

@misc{bi2025motusunifiedlatentaction,
            title={Motus: A Unified Latent Action World Model}, 
            author={Hongzhe Bi and Hengkai Tan and Shenghao Xie and Zeyuan Wang and Shuhe Huang and Haitian Liu and Ruowen Zhao and Yao Feng and Chendong Xiang and Yinze Rong and Hongyan Zhao and Hanyu Liu and Zhizhong Su and Lei Ma and Hang Su and Jun Zhu},
            year={2025},
            eprint={2512.13030},
            archivePrefix={arXiv},
            primaryClass={cs.CV},
            url={https://arxiv.org/abs/2512.13030}, 
}

motus-robotics/Motus_robotwin2

Motus: RoboTwin 2.0 Fine-Tuned Checkpoint

Table of Contents