SoFar: Language-Grounded Orientation Bridges Spatial Reasoning and Object Manipulation

Zekun Qi13*, Wenyao Zhang23*, Yufei Ding34*, Runpei Dong5, Xinqiang Yu3
Jingwen Li4, Lingyun Xu4, Baoyu Li5, Xialin He5, Guofan Fan1, Jiazhao Zhang3, Jiawei He3
Jiayuan Gu6, Xin Jin7, Kaisheng Ma1, Zhizheng Zhang3†, He Wang34†, Li Yi189†
* equal contribution † corresponding authors
1Tsinghua University 2Shanghai Jiao Tong University 3Galbot 4Peking University 5UIUC
6ShanghaiTech University 7Eastern Institute of Technology 8Shanghai Qi Zhi Institute 9Shanghai AI Laboratory
Paper arXiv Video Code Huggingface

Highlight

1. Complex robotic manipulation tasks are constrained by the understanding of orientation, such as "upright a tilted wine glass", or "plugging a cord into a power strip."

2. We introduce the concept of semantic orientation, representing the object orientation condition on open vocabulary language. Such as the orientation of "top," "handle," and "pouring water."

3. We construct OrienText300K, a large paired dataset of point clouds, text, and orientation. We trained PointSO, the first Open-Vocabulary Orientation Model.

4. Based on PointSO, we propose SoFar, the first 6-DoF spatial understanding LLM, which achieves a 13.1% performance improvement on the 6-DoF object rearrangement task and a 47.2% improvement over OpenVLA on the SimplerEnv benchmark.

5. We propose two benchmarks, Open6DOR V2 and 6-DoF SpatialBench, which evaluate 6-DoF rearrangement capability and 6-DoF spatial understanding capability, respectively.

SoFar Robotic Manipulation Pipeline

Interpolation end reference image.

Given the language instruction, SoFar prompts the VLM to obtain task-oriented object phases and semantic orientation descriptions. Then, SoFar leverage foundation models Florence-2 and SAM to segment depth point clouds and our PointSO to obtain semantic orientations. Summarizing 3D object-centric information, an orientation-aware scene graph is constructed and encoded into languages. The VLM takes the RGB image and the scene graph and outputs the queried spatial understanding VQA or translation for manipulation.

Real-World Experiments

We show the quantitative evaluation of zero-shot real world language-grounded rearrangement with SoFar. We design 60 diverse real world experimental tasks involving over 100 diverse objects.

Interpolation end reference image.

Demo

SoFar is capable of performing various complex object manipulation with spatial relationships and re-orientation tasks and can generalize across different embodiments, such as dexterous hands.

Pick up the teapot and pour water into the cup.
Take out the test tube with the green solution.
Rotate Loopy to face the yellow dragon doll.
Place the right bottle into the box and arrange it in a 3×3 pattern.
Rotate the flashlight to illuminate the loopy.
Upright the bottle.
Upside down the bottle.
Put the chili into the basket.
Upright the fallen wine glass and arrange it neatly in a row with the other wine glasses.
Insert the pen into the pen holder.
Pick the highest box and place it on the right.
Pick the box and place it to the right of the doll.
pick baseball and place it in the cart, then turn the cart to right.
Pull out a tissue.
Pick up the cabbage and place it in the basket.
Pour out chips from the chips cylinder to the plate.
Aim the camera at the toy truck.
Pick up the Lego blocks and place it between the two toy truck.

Navigation Demo

Semantic orientation can not only be applied to manipulation tasks but also to robotic navigation task. This orientation-aware constraint enhances the navigation process by ensuring precise alignment with the desired orientation, thereby improving task performance in scenarios where directionality is critical.

Move to facing the front of the microwave.
Move to facing the third chair‘s back.

Long Horizon Demo

Our model can complete multiple consecutive tasks, including pick & place, articulated object manipulation, and 6-DoF object rearrangement.

Clean the table.
6-DoF Shelf Rearrangement.

Close-Loop Planning

We demonstrate the closed-loop replan capabilities of SoFar within Simpler-Env.
In (a), model accidentally knocks over the Coke can during motion. Subsequently, we re-plan and successfully achieve the grasp.
In (b), model initially misidentified the coke can as a Fanta can. After correction, the model re-identifies and locates the correct object.

(a) Pick coke can.
(b) Pick coke can.

BibTeX

@article{qi2025sofar,
      author = {Qi, Zekun and Zhang, Wenyao and Ding, Yufei and Dong, Runpei and Yu, Xinqiang and Li, Jingwen and Xu, Lingyun and Li, Baoyu and He, Xialin and Fan, Guofan and Zhang, Jiazhao and He, Jiawei and Gu, Jiayuan and Jin, Xin and Ma, Kaisheng and Zhang, Zhizheng and Wang, He and Yi, Li},
      title = {SoFar: Language-Grounded Orientation Bridges Spatial Reasoning and Object Manipulation},
      journal = {arXiv preprint arXiv:2502.13143},
      year = {2025}
    }