Xin Cheng, Wangding Zeng, Damai Dai, Qinyu Chen, Bingxuan Wang, Zhenda Xie, Kezhao Huang, Xingkai Yu, Zhewen Hao, Han Zhang, Yu-Kun Li, Huishuai Zhang, Dongyan Zhao, Wenfeng Liang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

We present JanusFlow, a powerful framework that unifies image understanding and generation in a single model. JanusFlow introduces a minimalist architecture that integrates autoregressive language models with rectified flow, a state-of-the-art method in generative modeling. Our key finding demonstrates that rectified flow can be straightforwardly trained within the large language model framework, eliminating the need for complex architectural modifications. To further improve the performance of our unified model, we adopt two key strategies: (i) decoupling the understanding and generation encoders, and (ii) aligning their representations during unified training. Extensive experiments show that JanusFlow achieves comparable or superior performance to specialized models in their respective domains, while significantly outperforming existing unified approaches. This work represents a step toward more efficient and versatile vision-language models.

We introduce Janus, an autoregressive framework that unifies multimodal understanding and generation. Prior research often relies on a single visual encoder for both tasks, such as Chameleon. However, due to the differing levels of information granularity required by multimodal understanding and generation, this approach can lead to suboptimal performance, particularly in multimodal understanding. To address this issue, we decouple visual encoding into separate pathways, while still leveraging a single, unified transformer architecture for processing. The decoupling not only alleviates the conflict between the visual encoder’s roles in understanding and generation, but also enhances the framework’s flexibility. For instance, both the multi-modal understanding and generation components can independently select their most suitable encoding methods. Experiments show that Janus surpasses previous unified model and matches or exceeds the performance of task-specific models. The simplicity, high flexibility, and effectiveness of Janus make it a strong candidate for next-generation unified multimodal models. The code will be made available.

Four degree-of-freedom selective compliance assembly robot arm (SCARA) robots are increasingly used in industry due to their unique advantages in speed and accuracy. When integrated into visual-guided systems with cameras, SCARA robots require hand-eye and robot-world (HE&RW) calibration to establish the system's geometric relationships. The conventional HE&RW calibration methods are pose-based or point-based, which have been widely validated on full degree-of-freedomarticulated robots. However, these methods may appear singular and unusable due to SCARA's restricted movement. Besides, some calibration methods designed for this situation still impose additional requirements or lack in-depth singularity analysis. Inspired by this, we conduct a thorough study on HE&RW calibration for SCARA robots. First, we analyze the reasons for the SCARA's singularity of conventional methods from the perspective of nonlinear least squares. Then, we redefine parameters with clear geometric interpretation and propose two nonsingular HE&RW calibration methods. Note that comparative studies of pose-based and point-based methods on both articulated and SCARA robots are carried out. Finally, the effectiveness, adaptability, and generality of the proposed methods are validated on both synthetic and real data.

Damai Dai, Chengqi Deng, Chenggang Zhao, R.x. Xu, Huazuo Gao, Deli Chen, Jiashi Li, Wangding Zeng, Xingkai Yu, Y. Wu, Zhenda Xie, Y.k. Li, Panpan Huang, Fuli Luo, Chong Ruan, Zhifang Sui, Wenfeng Liang. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

Hand-eye calibration aims to relate what the camera sees to where the robot moves, which is crucial for vision-guided robot systems and has received much attention in the robotics community. The classical hand-eye calibration flow estimates camera extrinsic poses first, then uses homogeneous pose alignment to solve the hand-eye calibration problem. Since this two-step procedure is trivial and prone to error propagation, point-alignment-based hand-eye calibration is currently gaining popularity. The point alignment approach is promising but still immature, as evidenced by the lack of a unified formulation for different calibration targets, the requirement of pose for initialization, and the reliance on an optimization solver and its numerical differentiation. These issues are addressed item by item in this article. We first formulate the hand-eye calibrations for multi-point, single-point, and patterned targets via 2D-3D generative point alignment. Then, we propose a generic initialization on a single-point sequence covering all the above target cases. Following that, we infer the analytical Jacobian matrix in detail and develop the sparsity for the pose-perturbation refinement. Finally, both numerical simulations and real-world experiments are provided to verify that our approach is more accurate, efficient, and robust than state-of-the-art methods. The codes and datasets are open-source and can be found at http://github.com/MatthewJin001/GPA-HEC.Note to Practitioners—Vision-guided robots have been widely deployed in flexible automation, and only through hand-eye information can visual perception be used to guide robotic movement. However, the classical hand-eye calibration methods perform the hand-eye parameter estimation with beforehand camera poses, which is time-consuming and prone to error propagation. This work investigates one-step hand-eye calibrations based on direct point alignment. We provide simultaneous hand-eye and target estimation methods for multi-point, single-point, and patterned calibration targets by solving 2D-3D generative point alignment problems. Generally speaking, the single-point method achieves the least time consumption, the patterned method achieves the highest accuracy with a precise pattern, and the multi-point method achieves the strongest robustness and adaptability. With reliable robot information, the proposed methods are more accurate even in pattern deformation cases compared with the classical two-step methods. With the analytical Jacobian refinement, our methods are solver-free, lightweight, and suitable for flexible deployment. Furthermore, without the range requirement of pose acquisition, our methods offer the ability to integrate path planning into autonomous hand-eye calibration.