ECCV 2026

UniPart

Towards Zero-shot Language-Grounded 3D Part Segmentation for Embodied Interaction

Xinqiang Yu^1,4,6,*, Zekun Qi^3,4,*, Jiawei He⁶, Wenyao Zhang⁵, Xuchuan Chen^3,4

Guocai Yao⁶, Li Yi^3,†, Zhaoxiang Zhang^1,†, He Wang^2,4,†

* equal contribution † corresponding authors

¹ NLPR, Institute of Automation, Chinese Academy of Sciences ² Peking University
³ Tsinghua University ⁴ Galbot ⁵ Shanghai Jiao Tong University ⁶ Beijing Academy of Artificial Intelligence

Paper Code (Coming Soon) Dataset (Coming Soon) BibTeX

UniPart overview: LangPart-1M data engine, UniPart model, and language-conditioned real-world part grasping

UniPart maps a free-form phrase to a functional 3D part mask. We scale supervision with LangPart-1M (160K+ Objaverse assets, 8M text–part pairs) and use the predicted masks for language-conditioned grasping.

Abstract

Fine-grained robotic manipulation depends on understanding parts, not only whole objects. Existing 3D foundation models tend to be either generalized but object-aware, or part-aware but limited to closed-set taxonomies, which weakens zero-shot transfer. We study text-conditioned 3D part segmentation, where a free-form phrase selects a functional part on a point cloud. We introduce UniPart, a feed-forward cross-modal 3D Transformer that conditions on a CLIP text embedding. To scale supervision, we build LangPart-1M with 160K+ Objaverse assets and 8M text-to-part pairs using multi-view consistent part generation. We further manually label a high-quality subset, LangPart-4K, for fine-tuning and evaluation. UniPart achieves strong zero-shot results on open-vocabulary part benchmarks and transfers to language-conditioned part grasping in the real world.

Keywords: Zero-Shot · 3D Part Segmentation · Embodied Interaction

Highlights

160K+3D objects

8Mtext–part pairs

0.4sfeed-forward inference

90.5%real-world part accuracy

LangPart-1M and LangPart-1K

A large-scale dataset with 160K objects and 8M text-to-part pairs, plus a 1K-object open-vocabulary benchmark with more than 4K annotations.

Scalable data engine

Multi-view rendering, SAM proposals, Set-of-Mark prompting, and text-guided merging lift 2D parts into consistent 3D masks. LangPart-4K is verified by ten annotators.

A simple cross-modal Transformer

UniPart injects a frozen CLIP text embedding into every layer by addition, keeping the model lightweight while remaining strongly language-conditioned.

Embodied validation

On real sensor point clouds, UniPart produces part masks that can be passed to GraspNet and DexGraspNet 2.0 for language-conditioned grasping.

Method

Existing 3D models often fall into two patterns that leave a gap for dexterous manipulation: promptable but coarse instance separation, or part-aware but closed-set taxonomies. UniPart treats language-grounded part segmentation as the missing interface: take a natural language instruction and return the corresponding 3D part mask.

LangPart-1M annotation pipeline — LangPart-1M generation: query a part list, segment twelve rendered views with SAM and Set-of-Mark, merge regions with text guidance, then lift aligned masks to the point cloud.

UniPart three-stage training — Three-stage curriculum. Stage 1 aligns point features with CLIP image patches. Stages 2–3 pretrain on LangPart-1M and fine-tune on manually verified LangPart-4K, injecting CLIP text embeddings layer-wise.

Dataset and Qualitative Results

LangPart qualitative examples and dataset scale — Open-vocabulary part examples from LangPart, compared with FIND3D in object and part scale. The dataset covers both single-view and complete object point clouds.

Results

Zero-shot mIoU on Objaverse-General, following the FIND3D protocol with two prompt formats.

Method	Seen categories		Unseen categories
Method	{part} of a {object}	{part}	{part} of a {object}	{part}
PointCLIPV2	9.81	11.27	10.27	11.09
PartSLIP++	2.69	15.03	0.57	10.43
OpenMask3D	11.81	11.93	7.01	10.31
FIND3D	33.78	34.10	26.21	27.41
UniPart (ours)	46.43	49.27	41.46	45.23

UniPart is also best on ShapeNetPart, PartNet-E, and our LangPart-1K split. A controlled study shows both the dataset and the architecture contribute: FIND3D on LangPart-1M reaches 38.42 mIoU, while UniPart on LangPart-1M reaches 47.25.

Real-World Part Grasping

We run language-conditioned grasping on a Franka Panda with a gripper and a UR arm with a Shadow Hand. A single RealSense D415 provides the observation. UniPart predicts the target part, which is then passed to GraspNet or DexGraspNet 2.0.

Real-world Franka and UR Shadow Hand setup — Real-world setup. On 20 objects, UniPart reaches 90.5% part segmentation accuracy and 85.0% grasp success on the intended part.

BibTeX

@inproceedings{yu2026unipart,
  title     = {UniPart: Towards Zero-shot Language-Grounded 3D Part Segmentation
               for Embodied Interaction},
  author    = {Xinqiang Yu and Zekun Qi and Jiawei He and Wenyao Zhang and
               Xuchuan Chen and Guocai Yao and Li Yi and Zhaoxiang Zhang and He Wang},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year      = {2026}
}