Pelican-Unified 1.0: A Unified Embodied Intelligence Model (UEI) for Understanding, Reasoning, Imagination and Action Embodied IntelligenceWorld ModelsUnified Models
arXiv Technical Report | May 14, 2026
Beijing Innovation Center of Humanoid Robotics (X-Humanoid), WFM System Group; Yinda Chen (core contributor, ranked second)

Pelican-Unified 1.0, also reported as Pelican-Unify 1.0, unifies understanding, reasoning, future imagination, and action in one embodied intelligence loop. It uses a single VLM for scene/instruction understanding and task-oriented reasoning, plus a Unified Future Generator that jointly predicts future videos and actions in the same denoising process.

It ranks first on WorldArena with 66.03 EWM, reaches 98.12% 3D accuracy, achieves 93.5 average success on RoboTwin, and scores 64.7 across eight VLM benchmarks among comparable-scale models. The model has been validated on UR5e arms and the Tiangong humanoid robot for zero-shot long-horizon tasks such as interface insertion, waterproofing, and object manipulation.

TokenUnify: Scaling Up Autoregressive Pretraining for Computer Vision Computer VisionSelf-Supervised Learning
ICCV | October 25, 2025
Yinda Chen*; Haoyuan Shi*; Xiaoyu Liu; Te Shi; Ruobing Zhang; Dong Liu; Zhiwei Xiong; Feng Wu

Code

Dataset

Weights

TokenUnify proposes a hierarchical predictive coding framework for computer vision, reducing autoregressive error from O(K) to O(√K). It introduces a dataset with 1.2 billion annotated voxels and achieves 44% improvement over training from scratch.

MaskTwins: Dual-form Complementary Masking for Domain-Adaptive Image Segmentation Domain AdaptationPretraining Methods
ICML | July 13, 2025
Jiawen Wang; Yinda Chen* (Theory Contribution & Project Leader); Xiaoyu Liu; Che Liu; Dong Liu; Jianqing Gao; Zhiwei Xiong

Code

Poster

MaskTwins introduces a dual-form complementary masking strategy for domain-adaptive image segmentation, effectively bridging the domain gap through coordinated spatial and feature-level masking mechanisms.

Learning multiscale consistency for self-supervised electron microscopy instance segmentation Computer VisionPretraining Methods
ICASSP | April 13, 2024
Yinda Chen; Wei Huang; Xiaoyu Liu; Shiyu Deng; Qi Chen; Zhiwei Xiong

Code

A pretraining framework for volume instance segmentation is proposed. It enforces multiscale consistency and shows good performance in instance segmentation tasks.

Self-Supervised Computer Vision with Multi-Agent Reinforcement Learning Computer VisionSelf-Supervised Learning
IJCAI (oral) | August 17, 2023
Yinda Chen; Wei Huang; Shenglong Zhou; Qi Chen; Zhiwei Xiong

Code

Pretrain Data

CREMI

VNC

This paper proposes a decision-based MIM for computer vision segmentation. It uses MARL to optimize masking, outperforming alternatives.

GTGM: Generative Text-Guided 3D Vision-Language Pretraining for Medical Image Segmentation Vision-LanguageMedical Imaging
ICCV Workshop | October 25, 2025
Yinda Chen*; Che Liu*; Wei Huang; Xiaoyu Liu; Haoyuan Shi; Sibo Cheng; Rossella Arcucci; Zhiwei Xiong

Code

GTGM extends Vision-Language Pretraining to 3D medical images by leveraging LLMs to generate synthetic textual descriptions, enabling text-guided representation learning without paired medical text. Combined with a negative-free contrastive learning strategy, GTGM achieves state-of-the-art performance across 10 CT/MRI segmentation datasets.

EMPOWER: Evolutionary Medical Prompt Optimization With Reinforcement Learning Vision-LanguageMultimodal Learning
IEEE Journal of Biomedical and Health Informatics | October 16, 2025
Yinda Chen*; Yangfan He*; Jing Yang; Dapeng Zhang; Zhenlong Yuan; Muhammad Attique Khan; Jamel Baili; Por Lip Yee

EMPOWER proposes an evolutionary framework for prompt optimization through specialized representation learning and multi-dimensional evaluation. It achieves 24.7% reduction in factual errors and 15.3% higher preference scores.

BIMCV-R: A Landmark Dataset for 3D CT Text-Image Retrieval Vision-LanguageMultimodal Learning
MICCAI | October 06, 2024
Yinda Chen; Che Liu; Xiaoyu Liu; Rossella Arcucci; Zhiwei Xiong

Dataset

This paper presents BIMCV-R, a 3D CT text-image retrieval dataset, and MedFinder. Tests show MedFinder outperforms baselines in related tasks.

Unsupervised Domain Adaptation for EM Image Denoising with Invertible Networks Domain AdaptationImage Denoising
IEEE Transactions on Medical Imaging | July 29, 2024
Shiyu Deng; Yinda Chen; Wei Huang; Ruobing Zhang; Zhiwei Xiong

Code

The paper proposes an unsupervised domain adaptation method for EM image denoising with invertible networks, outperforming existing methods.

Learned Image Coding with Generative Reference of Conditional Latents Image CompressionGenerative Models
IEEE Transactions on Pattern Analysis and Machine Intelligence | Accepted, 2025
Siqi Wu*; Yinda Chen*; Weiming Chen; Dong Liu; K. C. Ho; Zhihai He

Code

GRCL presents a generic framework that exploits semantically correlated external images as conditional coding references in the latent domain. Three reference generation methods are investigated: local dictionary retrieval, web-based image search, and diffusion-based image-text-image generation. Theoretical analysis proves robustness to reference perturbations via subspace recovery error bounds. Achieves up to 1.5 dB PSNR gain over state-of-the-art methods with only ~0.005 bpp overhead.

Condition-generation Latent Coding with an External Dictionary for Deep Image Compression Image Compression
AAAI (oral) | March 06, 2025
Siqi Wu; Yinda Chen*; Dong Liu; Zhihai He

Code

Weights

The paper proposes CLC for deep image compression. It uses a dictionary to generate references, shows good performance, and has theoretical analysis.

MaskFactory: Towards High-quality Synthetic Data Generation for Dichotomous Image Segmentation Multimodal Learning
NeurIPS | October 17, 2024
Haotian Qian; Yinda Chen*; Shengtao Lou; Fahad Shahbaz Khan; Xiaogang Jin; Deng-Ping Fan

Project

Code

MaskFactory proposes a two-stage method to generate high-quality synthetic datasets for DIS, outperforming existing methods in quality and efficiency.