UNITAS

UNITAS is a 3D-native world action model for embodied manipulation. This repository provides inference checkpoints.

Paper · Project · Code

RoboTwin · LIBERO · VLABench · Scene flow

RoboTwin

Policies for RoboTwin Clean and Randomized evaluation. Use --domain robotwin.

Training	Checkpoint	Download
From scratch	unitas_robotwin_scratch.pt	Download
Pretrained, then fine-tuned	unitas_robotwin_finetuned.pt	Download

LIBERO

Policies for LIBERO; the fine-tuned model also supports LIBERO-Plus evaluation. Use --domain libero.

Training	Checkpoint	Download
From scratch	unitas_libero_scratch.pt	Download
Pretrained, then fine-tuned	unitas_libero_finetuned.pt	Download

VLABench

Policies for VLABench with the five custom evaluation tracks. Use --domain vlabench.

Training	Checkpoint	Download
From scratch	unitas_vlabench_scratch.pt	Download
Pretrained, then fine-tuned	unitas_vlabench_finetuned.pt	Download

Scene flow

Jointly trained on RoboTwin, LIBERO and VLABench for action and dense scene flow prediction. The model predicts gripper motion once, then uses it to condition scene flow prediction.

Training	Checkpoint	Download
Joint training	unitas_scene_flow.pt	Download

Usage

Install the inference code from UNITAS and run the commands below from the repository root. Each .pt file contains model tensors and inference configuration (model and args), without optimizer state or training metadata.

Download Florence-2-large (shared by all UNITAS checkpoints) and the checkpoint you want to use:

python -m pip install "huggingface_hub[cli]>=0.34,<1"
hf download microsoft/Florence-2-large \
  --local-dir checkpoints/Florence-2-large --exclude "*.bin"
hf download RobbinWang123/Unitas unitas_robotwin_finetuned.pt --local-dir checkpoints

Start a policy server. Set the domain and checkpoint to match the benchmark above; both scratch and fine-tuned models use the same interface.

python src/unitas/inference/server.py --domain robotwin \
  --checkpoint checkpoints/unitas_robotwin_finetuned.pt \
  --florence-dir checkpoints/Florence-2-large --port 8000

The benchmark client connects to http://127.0.0.1:8000.

For scene flow, download unitas_scene_flow.pt and run a bundled sample from the code checkout:

python src/unitas/scene_flow/inference.py \
  --checkpoint checkpoints/unitas_scene_flow.pt \
  --florence-dir checkpoints/Florence-2-large \
  --input assets/scene_flow/observations/robotwin_7c59d8eb33_0088.npz \
  --output outputs/scene_flow.npz
python src/unitas/scene_flow/visualize.py --input outputs/scene_flow.npz --port 8080

Open http://127.0.0.1:8080 for the interactive visualization.

Citation

If you find UNITAS useful, please cite:

@misc{wang2026unitas3dnativeworldaction,
  title={UNITAS: A 3D-Native World Action Model for Embodied Manipulation},
  author={Ruixiang Wang and Yongyi Su and Wenlve Zhou and Bo Yue and Hengyan Liu and Dekun Lu and Yuxin Tian and Yihan Fang and Zerui Wu and Xing Hu and Jietao Chen and Yong Guo and Ziyan He and Junbin Yuan and Guiliang Liu and Xiaofen Xing and Kui Jia},
  year={2026},
  eprint={2610.12099},
  archivePrefix={arXiv},
  primaryClass={cs.RO},
  url={https://arxiv.org/abs/2610.12099},
}