RoboBrain 2.5: Depth in Sight, Time in Mind.

BAAI RoboBrain Team

Core Contributor, Technical Report 2026

Project | Paper | Code | Checkpoints

RoboBrain 2.5 upgrades embodied intelligence along two key axes: depth-aware 3D spatial reasoning and dense temporal value estimation. It predicts physically grounded 3D manipulation traces and step-aware progress signals across viewpoints, improving fine-grained manipulation planning and execution feedback.

RoboBrain 2.0: See Better. Think Harder. Do Smarter.

BAAI RoboBrain Team

First Author, Core Contributor, Technical Report 2025

Project | Paper | Code | Checkpoints

RoboBrain 2.0 is an embodied vision-language foundation model family with 7B and 32B variants, designed to unify perception, reasoning, and planning for complex physical-world tasks. The 32B model reports leading performance on both spatial and temporal embodied benchmarks, with strong capabilities in affordance prediction, spatial referring, trajectory forecasting, closed-loop interaction, and multi-agent long-horizon planning.

RoboBrain: A Unified Brain Model for Robotic Manipulation from Abstract to Concrete

Yuheng Ji*, Huajie Tan*, Jiayu Shi*, Xiaoshuai Hao*, Yuan Zhang, Hengyuan Zhang, Pengwei Wang, Mengdi Zhao, Yao Mu, Pengju An, Xinda Xue, Qinghang Su, Huaihai Lyu, Xiaolong Zheng, Jiaming Liu, Zhongyuan Wang, Shanghang Zhang

First Author, CVPR 2025

Project | Paper | Code | Checkpoints | Datasets

We developed RoboBrain, a VLM-based model that combines robotic and general multi-modal data, utilizes a multi-stage training strategy, and incorporates long videos and high-resolution images to improve its robotic manipulation capabilities. Extensive experiments demonstrate that RoboBrain achieves SOTA performance across various robotic tasks, highlighting its potential to advance robotic brain capabilities.

Reason-RFT: Reinforcement Fine-Tuning for Visual Reasoning of Vision Language Models

Huajie Tan*, Yuheng Ji*, Xiaoshuai Hao*, Minglan Lin, Pengwei Wang, Shanghang Zhang

First Author, NeurIPS 2025

Project | Paper | Code | Checkpoints | Datasets

We developed Reason-RFT, a reinforcement fine-tuning framework that enhances visual reasoning capabilities in vision-language models. Reason-RFT employs a two-phase training strategy: (1) SFT with curated CoT data to activate reasoning potential, followed by (2) GRPO-based reinforcement learning to generate diverse reasoning-response pairs.

RoboOS-NeXT: A Unified Memory-based Framework for Lifelong, Scalable, and Robust Multi-Robot Collaboration

Huajie Tan*, Yuheng Ji*, Cheng Chi*, Xiansheng Chen*, Zhongxia Zhao, Xiaoshuai Hao, Yaoxu Lyu, Mingyu Cao, Junkai Zhao, Huaihai Lyu, Enshen Zhou, Ning Chen, Yankai Fu, Cheng Peng, Wei Guo, Dong Liang, Zhuo Chen, Mengsi Lyu, Chenrui He, Yulong Ao, Yonghua Lin, Pengwei Wang, Zhongyuan Wang, Shanghang Zhang

First Author, ArXiv 2025

Project | Paper | Code

We present RoboOS, a unified memory-based framework for multi-robot collaboration. At its core, the Spatio-Temporal–Embodiment Memory (STEM) integrates spatial, temporal, and embodiment information to support long-horizon learning, heterogeneous coordination, and fault recovery. Experiments in diverse tasks show RoboOS enables lifelong, scalable, and robust collaboration.

Robo-Dopamine: General Process Reward Modeling for High-Precision Robotic Manipulation

Huajie Tan*, Sixiang Chen*, Yijie Xu*, Zixiao Wang, Yuheng Ji, Cheng Chi, Yaoxu Lyu, Zhongxia Zhao, Xiansheng Chen, Peterson Co, Shaoxuan Xie, Guocai Yao, Pengwei Wang, Zhongyuan Wang, Shanghang Zhang

CVPR 2026

Project | Paper | Code | Checkpoints

Robo-Dopamine introduces a step-aware multi-view reward modeling framework (Dopamine-Reward) and a policy-invariant shaping strategy (Dopamine-RL) for robotic RL. The method improves reward assessment quality and substantially boosts online policy learning efficiency on both simulated and real-world manipulation tasks.

Action-Sketcher: From Reasoning to Action via Visual Sketches for Long-Horizon Robotic Manipulation

Huajie Tan*, Peterson Co*, Yijie Xu*, Shanyu Rong, Yuheng Ji, Cheng Chi, Xiansheng Chen, Qiongyu Zhang, Zhongxia Zhao, Pengwei Wang, Zhongyuan Wang, Shanghang Zhang

CVPR 2026

Project | Paper | Code

Action-Sketcher proposes a See-Think-Sketch-Act loop that inserts editable visual sketches as an intermediate representation between language reasoning and robot actions. This design improves grounding, long-horizon robustness, and interpretability in cluttered and dynamic manipulation environments.

VisualTrans: A Benchmark for Real-World Visual Transformation Reasoning

Yuheng Ji*, Yipu Wang*, Yuyang Liu, Xiaoshuai Hao, Yue Liu, Yuting Zhao, Huaihai Lyu, Xiaolong Zheng

First Author, ArXiv 2025

Paper | Code | Datasets

VisualTrans is the first real-world benchmark for Visual Transformation Reasoning (VTR), evaluating spatial, procedural and quantitative reasoning across 12 human-object interaction tasks. While current models perform well on static tasks, they show significant limitations in dynamic, multi-step reasoning, revealing critical gaps in temporal and causal understanding for intelligent systems.

MathSticks: A Benchmark for Visual Symbolic Compositional Reasoning with Matchstick Puzzles

Yuheng Ji*, Huajie Tan*, Cheng Chi*, Yijie Xu, Yuting Zhao, Enshen Zhou, Huaihai Lyu, Pengwei Wang, Zhongyuan Wang, Shanghang Zhang, Xiaolong Zheng

First Author, NeurIPS MATH-AI Workshop 2025

Paper | Code | Datasets

MathSticks is a benchmark for Visual Symbolic Compositional Reasoning (VSCR) that unifies visual perception, symbolic manipulation, and arithmetic consistency. Each task presents an incorrect matchstick equation in a seven-segment style. The goal is to move exactly one or two sticks—under strict stick-conservation and digit-legibility constraints—to make the equation mathematically correct.

Towards a Unified Understanding of Robot Manipulation: A Comprehensive Survey

Shuanghao Bai, Wenxuan Song, Jiayi Chen, Yuheng Ji, Zhide Zhong, Jin Ynag, Han Zhao, Wanqi Zhou, Wei Zhao, Zhe Li, Pengxiang Ding, Cheng Chi, Haoang Li, Chang Xu, Xiaolong Zheng, Donglin Wang, Shanghang Zhang, Badong Chen

ArXiv 2025

Paper | Repo

Synthesizing over 1200+ publications, this survey restructures the landscape of robotic manipulation with a unified taxonomy for planning and control. We also provide the first systematic dissection of the key bottlenecks—data, utilization, and generalization—poised to define the next era of progress.

ManipLVM-R1: Reinforcement Learning for Reasoning in Embodied Manipulation with Large Vision-Language Models

Zirui Song*, Guangxian Ouyang*, Mingzhe Li, Yuheng Ji, Chenxi Wang, Zixiang Xu, Zeyu Zhang, Xiaoqing Zhang, Qian Jiang, Zhenhao Chen, Zhongzhi Li, Rui Yan, Xiuying Chen

AAAI 2026

Paper

VLMs enhance robotic manipulation but rely on costly annotated data, limiting OOD adaptability. We propose ManipLVM-R1, a RL framework with verifiable rewards (RLVR), replacing supervision to optimize task outcomes for better generalization. Two rule-based rewards drive physical reasoning, achieving strong performance on fewer data and OOD scenarios.

EgoPrompt: Prompt Learning for Egocentric Action Recognition

Huaihai Lyu, Chaofan Chen, Yuheng Ji, Changsheng Xu

ACM MM 2025

Paper

EgoPrompt is a prompt-learning framework for egocentric action recognition that jointly models verbs and nouns by capturing their semantic relationships. It introduces a Unified Prompt Pool and a Diverse Pool Criteria to encourage rich, disentangled representations. EgoPrompt achieves state-of-the-art performance on Ego4D, EPIC-Kitchens, and EGTEA across various generalization benchmarks.

Alleviating Performance Disparity in Adversarial Spatiotemporal Graph Learning under Zero-inflated Distribution

Songran Bai, Yuheng Ji, Yue Liu, Xingwei Zhang, Xiaolong Zheng, Daniel Dajun Zeng

AAAI 2025 (Oral)

Paper

Spatiotemporal graph learning under zero-inflated distribution is vital for urban risk management but is susceptible to adversarial attacks. Traditional adversarial training increases performance disparities between classes. We propose the MinGRE framework to reduce these disparities and enhance robustness.

Enhancing Adversarial Robustness of Vision-Language Models through Low-Rank Adaptation

Yuheng Ji*, Yue Liu*, Zhicheng Zhang, Zhao Zhang, Yuting Zhao, Xiaoshuai Hao, Gang Zhou, Xingwei Zhang, Xiaolong Zheng

First Author, ICMR 2025

Paper

We propose a parameter-efficient adversarial adaptation method named AdvLoRA by low-rank adaptation to improve the robustness of vision-language models.

FastRSR: Efficient and Accurate Road Surface Reconstruction from Bird’s Eye View

Yuting Zhao*, Yuheng Ji*, Xiaoshuai Hao, Shuxiao Li

First Author, ACM MM 2025

Paper

Road surface reconstruction is crucial for autonomous driving, enabling the understanding of road surface conditions. We present two BEV-based RSR models: FastRSR-mono and FastRSR-stereo, offering superior efficiency and accuracy across elevation error and processing speed.

What Really Matters for Robust Multi-Sensor HD Map Construction?

Xiaoshuai Hao, Yuting Zhao, Yuheng Ji, Luanyuan Dai, Shuai Cheng, Rong Yin

IROS 2025 (Oral)

Paper

This paper enhances HD map construction robustness via data augmentation, a new fusion module, and modality dropout. It improves performance under sensor corruptions and achieves SOTA accuracy on NuScenes.

MSC-Bench: Benchmarking and Analyzing Multi-Sensor Corruption for Driving Perception

Xiaoshuai Hao, Guanqun Liu, Yuting Zhao, Yuheng Ji, Mengchuan Wei, Haimei Zhao, Lingdong Kong, Rong Yin, Yu Liu

ICME 2025 (Oral)

Project | Paper

This work introduces the Multi-Sensor Corruption Benchmark (MSC-Bench), the first comprehensive benchmark aimed at evaluating the robustness of multi-sensor autonomous driving perception models against various sensor corruptions.

Learning Hash Subspace from Large-Scale Multi-modal Pre-Training: A CLIP-Based Cross-modal Hashing Framework

Yuheng Ji*, Xingwei Zhang*, Gang Zhou, Xiaolong Zheng, Daniel Dajun Zeng

First Author, The 11st C2 China 2023 (Outstanding Paper Award)

Paper

We propose a cross-modal hashing framework called CCMH (CLIP-based Cross-Modal Hashing), which facilitates the transferability of a well-trained real-value semantic subspace to a hash semantic subspace.

Embodied Robot Manipulation in the Era of Foundation Models: Planning and Learning Perspectives

Shuanghao Bai, Wenxuan Song, Jiayi Chen, Yuheng Ji, Zhide Zhong, Jin Yang, Han Zhao, Wanqi Zhou, Zhe Li, Pengxiang Ding, Cheng Chi, Chang Xu, Xiaolong Zheng, Donglin Wang, Haoang Li, Shanghang Zhang, Badong Chen

ArXiv 2025

Paper

This survey revisits robot manipulation in the foundation-model era through a unified two-level abstraction: high-level planning and low-level control. It organizes algorithmic design choices, summarizes representative methods, and highlights open challenges in scalability, data efficiency, multimodal physical interaction, and safety.

Towards Cross-View Point Correspondence in Vision-Language Models

Yipu Wang*, Yuheng Ji*, Yuyang Liu, Enshen Zhou, Ziqiang Yang, Yuxuan Tian, Ziheng Qin, Yue Liu, Huajie Tan, Cheng Chi, Zhiyuan Ma, Daniel Dajun Zeng, Xiaolong Zheng

First Author, ArXiv 2025

Project | Paper

This work introduces the Cross-View Point Correspondence task and CrossPoint-Bench to evaluate point-level cross-view grounding in VLMs. It further proposes CrossPoint-378K and CroPond to improve fine-grained correspondence accuracy for embodied affordance-centric interaction.

RoboTracer: Mastering Spatial Trace with Reasoning in Vision-Language Models for Robotics

Enshen Zhou, Cheng Chi, Yibo Li, Jingkun An, Jiayuan Zhang, Shanyu Rong, Yi Han, Yuheng Ji, Mengzhen Liu, Pengwei Wang, Zhongyuan Wang, Lu Sheng, Shanghang Zhang

ArXiv 2025

Project | Paper

RoboTracer targets metric-grounded multi-step spatial tracing for robotics by combining 3D-aware supervised fine-tuning with reinforcement fine-tuning under metric-sensitive process rewards. It improves spatial referring, measuring, and long-horizon trace generation in real-world embodied settings.

RoboMirror: Understand Before You Imitate for Video to Humanoid Locomotion

Zhe Li, Cheng Chi, Boan Zhu, Yangyang Wei, Shuanghao Bai, Yuheng Ji, Yibo Peng, Tao Huang, Pengwei Wang, Zhongyuan Wang, S.-H. Gary Chan, Chang Xu, Shanghang Zhang

ArXiv 2025

Paper

RoboMirror proposes a retargeting-free video-to-humanoid-locomotion pipeline under the principle of “understand before imitate”. By extracting visual motion intent from egocentric/third-person videos and conditioning diffusion policies, it narrows the gap between visual understanding and control.

Scaling Up AI-Generated Image Detection via Generator-Aware Prototypes

Ziheng Qin*, Yuheng Ji*, Renshuai Tao, Yuxuan Tian, Yuyang Liu, Yipu Wang, Xiaolong Zheng

First Author, CVPR 2026

Code | Paper

This work studies the scaling challenge of universal AI-generated image detection and proposes Generator-Aware Prototype Learning (GAPL). By combining canonical forgery prototypes with a two-stage low-rank adaptation strategy, it improves robustness across diverse generators.