GeoPredict: Leveraging Predictive Kinematics
and 3D Gaussian Geometry
for Precise VLA Manipulation

🔥 CVPR 2026 🔥

Jingjing Qian¹  Boyao Han²  Chen Shi¹  Lei Xiao¹  Long Yang¹  Shaoshuai Shi³  Li Jiang^1
1Chinese University of Hong Kong, Shenzhen   ²Hunan University   ³Voyager Research, Didi Chuxing

---

🔥 News

• [2026-07] We released the training code and dataset for GeoPredict.
• [2026-05] We released the inference code and checkpoints for GeoPredict.
• [2026-02] Our paper was accepted by CVPR2026 as a Highlight ! 🥳
• [2025-12] We released the paper and the project page for GeoPredict.

---

🎄 Overview

GeoPredict is a geometry-aware vision-language-action (VLA) framework for robotic manipulation. Existing methods are often limited by:

1. 2D-Centric Formulation: operate in 2D image space, lacking explicit 3D spatial modeling.
2. Reactive Control: map observations reactively, failing to anticipate future physical dynamics.
3. Geometric Inconsistency: view-independent predictions struggle to enforce 3D consistency.

GeoPredict addresses these limitations with:

1. Geometry-Aware VLA: augments VLA with predictive kinematic and 3D geometric priors.
2. Predictive 3D Modeling: forecasts workspace geometry using track-guided 3DGS refinement.
3. Lightweight Inference: uses predictive modules solely for training, reducing test-time overhead.

---

📝 TODO

• Release paper and project page.
• Release inference code and checkpoints.
• Release training code and dataset.

---

📚 Getting Started

1. Environment Setup & Inference
2. Training on RoboCasa

The training dataset is available on ModelScope.

---

📖 Framework

GeoPredict consists of three key components:

• (a) Trajectory-Level Kinematic Prediction: encodes motion history of robot keypoints into compact tokens via a Track Encoder, and predicts multi-step 3D keypoint trajectories using learnable future track queries.
• (b) Predictive 3D Gaussian Geometry: decodes a coarse 3D spatial query into initial Gaussian primitives to represent workspace geometry, and forecasts how the explicit 3D scene representation evolves across multiple future timesteps.
• (c) Track-Guided Refinement & Rendering: adaptively increases Gaussian density along predicted trajectories to capture task-relevant interaction regions, and supervises the predictive 3DGS exclusively through future depth-map rendering without color modeling.

---

📊 Results

We report strong performance on both RoboCasa Human-50 and LIBERO benchmarks, demonstrating the effectiveness of GeoPredict in geometry-intensive and spatially demanding manipulation tasks. Please see the paper for full tables, metrics, and more detailed analysis.

---

📬 Contact

If you have questions about the paper, feel free to open an issue or contact:

• Jingjing Qian: [email protected]

---

🔗 Citation

If you find our work helpful, please cite:

@misc{qian2025geopredict,
  title={GeoPredict: Leveraging Predictive Kinematics and 3D Gaussian Geometry for Precise VLA Manipulation},
  author={Jingjing Qian and Boyao Han and Chen Shi and Lei Xiao and Long Yang and Shaoshuai Shi and Li Jiang},
  year={2025},
  eprint={2512.16811},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2512.16811},
}