Authors丨Zhang Qiping Xing Lijuan
Editor丨Xing Lijuan
The competition among world models is shifting from 'generating the future' toward 'whether robots can really use them.'Videos can be generated ever more realistically, but once a world model enters a robot system, it must answer more concrete questions: can it predict state changes after an action is executed, can it support policy training, and can it ultimately complete tasks on real robots. WorldArena has pulled world models into the same 'examination hall.' This evaluation system was launched by a Tsinghua University team together with multiple universities at home and abroad, and is one of the most authoritative leaderboards in the world model field. WorldArena 1.0 had already extended world model evaluation from pure video generation to embodied tasks such as data generation, policy evaluation, and action planning. But overall it still focused mainly on visual input, offline tasks, and simulated environments. Systematic evaluation was still lacking for real robot execution, online reinforcement learning loops, visual-tactile multimodal perception, and the noise, latency, and error accumulation of real-world deployment.To address these gaps, WorldArena 2.0 has further upgraded the evaluation system.Modalities have expanded from pure vision to visual-tactile, functions from offline tasks to online reinforcement learning, and platforms from simulation environments to real robots. As one of the most influential international top conferences in robotics, IROS provided the stage where the WorldArena 2.0 Challenge built an arena for real-robot closed loops, with three tracks respectively targeting video quality, online reinforcement learning environments, and real robot manipulation. After so long a debate among the major world model schools, models under different technical routes have finally competed head-to-head in this arena.ShiQi Future, which explicitly follows the JEPA route,won the Track 1 championship,the spatial intelligence routewhile the BWM-Turbo model from Tongji University on took second place with a 0.84-point gap; meanwhile, also on the JEPA route, ChenHunXian Technologyranked second overall in Track 2. In addition, a notable detail: among the top two on all three tracks, academia and industry each hold one seat. The competition in the world model track has now seen a direct collision of academic and engineering capabilities. Regarding the event's positioning, multi-institution collaboration, track design, and leaderboard results, AI Technology Review also interviewed the Tsinghua team, organizers of the WorldArena 2.0 Challenge. Their answers further illustrate how this evaluation system is designed and coordinated as it moves toward online interaction and real-robot execution, and what its boundaries are.
01
WorldArena 2.0:What Makes the New Round of the 'Big Exam' Difficult
WorldArena 2.0 pushes world model evaluation further into online interaction and real robots.WorldArena 1.0 mainly examined two types of capabilities: one was the quality of future videos generated by world models, with 15 metrics covering visual quality (Visual Quality), motion quality (Motion Quality), content consistency (Content Consistency), physics adherence (Physics Adherence), 3D accuracy (3D Accuracy), and controllability (Controllability); the other looked at whether these predictions are actually useful for robots, including data generation, policy evaluation, and action planning. Evaluation of 14 representative models showed that video quality and embodied task performance do not always correspond. Generating sharper, smoother images does not necessarily mean the model performs better in data generation, policy evaluation, and action planning. WorldArena 1.0's coverage was also relatively limited:in terms of modality it only processed visual information, in terms of function it focused on offline tasks such as data generation, policy evaluation, and action planning, and in terms of platform it mainly stayed in simulation environments. Tactile feedback, online interaction, and execution performance on real robots were not covered by this evaluation.The upgrade of WorldArena 2.0 proceeds alongModality, Functionality, and Platform,three directions.In terms of modality, from pure vision to visual-tactile.It examines whether a model can process both visual frames and contact information.In terms of function, from offline evaluation to online reinforcement learning.It examines whether a world model can serve as an interactive environment to support policy training.In terms of platform, from mainly relying on simulation environments to including real robot evaluation.Beyond simulation environments such as RoboTwin 2.0 and LIBERO, WorldArena 2.0 further added real robot testing, bringing models' execution performance in physical environments into the evaluation. Based on this evaluation direction of WorldArena 2.0, the Challenge set up three tracks:Track 1 evaluates video quality, Track 2 evaluates the world model's capability as a reinforcement learning environment, and Track 3 evaluates real robot manipulation.Track 1 continues and upgrades the original video quality evaluation, while Track 2 and Track 3 are new tracks added in 2.0.
Track 1 set 70 official tasks. Of these, 50 are ID (In-Distribution) clean scenarios on a white tabletop, and the other 20 are OOD (Out-of-Distribution) scenarios with colored or textured tabletops. All competing models are tested on the same set of datasets.
This track examines video generation performance in long-horizon and out-of-distribution scenarios. OOD scenarios introduce visual conditions different from the ID scenarios by changing the tabletop's color and texture, in order to observe whether the model can continue to maintain object identity, spatial topology, and action causal consistency when facing distribution shifts. The final score is jointly determined by 15 metrics, covering image quality, motion quality, content consistency, interaction accuracy, 3D geometry, and instruction execution. JEPA Similarity is also included: V-JEPA extracts video features to compare how close the feature distribution of the generated video is to that of the real reference video.Track 2 moves further into continuous interaction.The world model needs to serve as a reinforcement learning environment, in which the policy repeatedly executes actions, obtains new states and rewards, and continues to update. After training, the policy must also return to RoboTwin 2.0 to test task success rate, verifying whether the world model truly supportson-policy training.This requires the world model to remain stable and action-controllable across multiple rounds of interaction, correctly reflect the state changes brought about by different actions, and generate reliable frames to provide the model with effective reward signals. A single prediction deviation can also be carried into the next round of interaction and keep accumulating over continuous interactions.
Track 3 puts the evaluation directly into real robot manipulation. The organizers also set up two types of tasks, purely visual and visuo-tactile: they include visual-driven operations such as wiping the table, pouring water, clearing the tabletop, and folding clothes, as well as tasks requiring contact information such as picking up a potato chip, peeling a cucumber, and plugging in a two-hole plug. On real hardware, contact force changes, sensor noise, and communication latency all directly affect task results. The focus is on examining the generalization robustness of the model and policy in real environments and their ability to self-recover after action deviations.From Track 1 to Track 3, the evaluation progressively advances from "can it predict," to "can it support learning," and finally to "can it complete tasks on a real robot."02
The three tracks have unveiled their leaderboards;in what ways is each champion team strong?
On September 16, the global final leaderboard of the WorldArena 2.0 Challenge was revealed. The three tracks were ranked and awarded independently; based on the prize amounts published on the official website, thetotal prize money is 7,000 US dollars.▎Track 1: WorldIncept, winning not just because the videos "look good"
A total of 81 models received final scores on the Track 1 final leaderboard.Visincept's WorldIncept ranked first with 72.33 points; Tongji University's School of Computer Science and Technology Spatial Intelligence team's BWM-Turbo ranked second with 71.49 points, a gap of 0.84 points between the two. Looking at the sub-scores, WorldIncept's advantage is not concentrated only on image quality. The final leaderboard results show thatWorldIncept ranked first on two dimensions, Physics Adherence and 3D Ranked first in both dimensions, including Accuracy (3D Accuracy)；Semantic Alignment scored 90.11, also first in that single metric.Physics Adherence concerns whether the generation process conforms to real-world motion and interaction laws; 3D Accuracy further examines depth, perspective, and spatial geometric relationships; semantic alignment looks at whether the generated results truly change according to task requirements. These leading metrics correspond to some extent with the accumulated expertise of the team behind WorldIncept. Visincept, the company behind WorldIncept, was incubated by IDEA Research Institute, has released the DINO-X series of vision models and the DINO-XGrasp universal grasping model, and partnered with Baidu Intelligent Cloud to launch the EgoTwin data engine. It relies on hundred-thousand-hour-scale Ego data and teleoperation data for robot dynamics modeling, and incorporates the team's accumulated expertise in general object understanding, spatial intelligence, and human/hand motion understanding. By contrast, Tongji University's Spatial Intelligence team focuses more on the stability of long-horizon generation. Built on the DiT architecture, BWM-Turbo adds designs such as first-frame guidance, dynamic memory, and action control, focusing on handling scene stability, action responsiveness, and temporal consistency issues in long-horizon generation. In this final leaderboard, BWM-Turbo ranked first on two metrics, Background Consistency and JEPA Similarity, which also reflects this solution's advantages in scene stability and representational consistency. The two solutions differ in technical orientation, but the focus of competition in Track 1 is already clear: video generation is only the foundation,whether physical consistency in spatial structure, action changes, and long-horizon rollouts can be stably maintained is becoming the key differentiator.▎Track 2: MW2, world models truly entering policy training
Track 2 is the online reinforcement learning track newly added in WorldArena 2.0, evaluatingwhether a world model can truly serve as an environment for training robot policies. The policy first produces actions, the world model predicts the future states after the actions, rewards are then computed based on these predictions and the policy is updated, and finally it is checked whether the trained policy can accomplish the task.The track is based on the RoboTwin environment and uses the Adjust Bottle task. This task requires the robot to re-adjust the bottle's pose and position, testing fine pose control and grasping stability.
On the final leaderboard,MW2 from Beijing Zhongguancun Academy ranked first with 72.93 points, TTWM from Morning-Evening Terminator Technology took second place with 72.40 points — a gap of only 0.53 points — and they were the only two of the 47 participating models to exceed 72 points. Beijing Zhongguancun Academy, behind MW2, has in recent years continuously invested in embodied intelligence and physical intelligence. TTWM, close behind, adopted a more concrete optimization scheme for world model training. It isthe competition entry version of Morning-Evening Terminator's goal-conditioned causal world model GCWM, focusing on modeling "how actions will change future states". Conditioned on the current state, the task goal and the robot's action, the model predicts the physical interactions and state changes after the action is executed. In the continuous rollouts of Track 2, the team also carried out targeted optimizations around goal regions, spatial geometry and cross-time consistency, and reduced error accumulation in multi-step prediction through designs such as frame-level noise and cold-start randomization. These designs are further reflected in downstream policy performance. Baseline VLA π0.5 achieved a success rate of 55.46%; after policy optimization using GCWM as the reinforcement learning environment, the VLA model's success rate rose to 72.40%. Track 2 pushes the focus of competition further toward policy learning — whether world models can truly improve downstream policy performance is the core of this track.▎Track 3: ViTacX, First Overall on Real-Robot Tasks
Track 3 is the real-robot evaluation track newly added in WorldArena 2.0, covering the AgileX dual-arm robot and the Franka Panda single-arm robot. Track 3.1 evaluates visuo-tactile manipulation, completing on AgileXclamping a potato chip, peeling a cucumber and plugging a two-hole plug — 3 tasks; Track 3.2 evaluates pure vision-based manipulation, with AgileX coveringwiping the table, pouring water, cleaning the tabletop, cleaning the tabletop on command, folding clothes and folding paper boxes — 6 tasks, while Franka also covers wiping the table, pouring water and cleaning the tabletop.Each task is evaluated by the success rate of execution on real robots; the overall leaderboard combines task difficulty and aggregates scores with weights for the vision and visuo-tactile directions. The robot must complete the operation within the specified number of steps without triggering a safety intervention.
On the overall leaderboard,ViTacX submitted by the Xiaomi Robotics MiRobot team ranked first with 76.64 points, while the OmniFlow team from ShanghaiTech University took second place with 67.37 points. ViTacX's main advantage comes from the pure vision sub-leaderboard. In Track 3.2 it ranked first with 85.00 points, achieving the highest individual scores on 4 tasks — wiping the table, pouring water, folding clothes and folding paper boxes — with the first three all scoring 100. In terms of the team's technical background, Xiaomi Robotics has long focused on embodied foundation models, VLA and real-robot execution, with research spanning visual perception, action generation, and continuous execution on real robots. These directions correspond exactly to the vision-driven manipulation and real-robot execution abilities examined in Track 3.2.Track 3.2 Pure Vision Manipulation LeaderboardOn the Track 3.1 visuo-tactile sub-leaderboard, however, ViTacX ranked third with 66.67 points. By contrast, OmniFlow held second place on both the pure vision and visuo-tactile sub-leaderboards. ShanghaiTech University behind it has long carried out research related to robot manipulation and tactile sensing, with embodied manipulation research at its Automation and Robotics Center covering tactile perception, teleoperation, imitation learning, reinforcement learning and multi-contact locomotion.Track 3.1 Visuo-Tactile Manipulation LeaderboardJudging from the two sub-leaderboards, ViTacX's overall advantage mainly comes from vision-driven real-robot manipulation, while its performance on visuo-tactile tasks is relatively lower.03
Under the same evaluation system,world model technical approaches each have their own focus
Judging from the publicly disclosed schemes and leaderboard results, WorldArena 2.0 has put different world model technical approaches into the same evaluation system for comparison, and what each scheme excels at is beginning to become clearer. The top two of Track 1 point to two different technical routes. WorldIncept's backer, VisionFuture, unequivocally takes JEPA latent-space route, while BWM-Turbo takes thespatial intelligence route. Models from both routes compete on the same stage: WorldIncept leads in physical plausibility and 3D accuracy, while BWM-Turbo ranks first in background consistency and JEPA representation similarity. Track 2 directly tests whether a world model can reliably serve as a policy training environment. The champion MW2 comes from the Beijing Zhongguancun Academy; the team focuses on robots' understanding of physical laws and combines this understanding with action learning and execution. The runner-up TTWM from the Twilight Line (Chenhunxian) Technology places more emphasis on action causality prediction, attending to target regions, spatial geometric constraints, object consistency across time, and error accumulation in continuous rollouts. In Track 3, ViTacX from Xiaomi's MiRobot robot team took first place overall, but its advantage in purely visual tasks did not fully carry over to vision-tactile manipulation. This shows that real robot manipulation requires not only seeing the environment clearly, but also processing contact information.whether multimodal information can stably enter the closed loop of action planning and controlis the harder part. The top teams on the three tracks each have their own strengths, showing unique advantages at different capability levels. Since participation is selective, whether any model can simultaneously reach leading levels in video generation, policy training, and real robot operation remains to be verified, and is worth looking forward to。04
Dialogue with the organizers:How the world model's "exam hall" was built
From the unified judgment of competition results across multiple institutions to how the evaluation setup reduces bias from single metrics, there are many technical and organizational questions behind the WorldArena 2.0 evaluation. The organizer, the Tsinghua team, also answered these questions in an interview. The following is the dialogue transcript, edited by AI Technology Review without altering the original meaning.WorldArena's positioning and evaluation mechanism
▎AI Technology Review: WorldArena has now become one of the core evaluation benchmarks for world models. How do you view this recognition? And where do you think this recognition mainly comes from?
Organizers:This recognition mainly comes from the shared needs of academia and industry:there is a need for an evaluation method to judge whether the future predicted by a world model can truly help robots execute actions. For embodied intelligence, one must look at whether the model can accurately respond to action commands, whether the predicted motion conforms to physical laws, and whether it can support policy learning and real execution. These capabilities are hard to judge just by watching videos and require dedicated tests. WorldArena 1.0 turned these questions into test procedures that can be executed under a unified standard, and 2.0 added online reinforcement learning and real robot interaction to furtherexamine simulation-to-real (Sim2Real) transfer and closed-loop feedback. For us, whether a benchmark has long-term value depends on whether it can help researchers and practitioners find the essential problems of models, explain their strengths and weaknesses, and improve methods accordingly.▎AI Technology Review: WorldArena involves multiple institutions including Tsinghua, Peking University, CMU, Stanford, Princeton, the University of Hong Kong, NUS, and the Institute of Automation of the Chinese Academy of Sciences. As the organizer, what role does the Tsinghua team specifically play?
Organizers: The Tsinghua team is mainly responsible for the research design and development of the evaluation framework, as well as competition evaluation and open-source community maintenance.This benchmark covers everything from video generation to real-robot operation, requiring substantial investment in software and hardware, computing power, and human resources; it is hard for a single team to accomplish, so collaboration among multiple institutions is essential. The specific work includesbuilding the evaluation pipeline、standardizing the handling of submitted code、connecting to cross-region remote inference services、maintaining the distributed evaluation cluster, as well asverifying and aggregating results. The real-robot track also requires coordinating with robot platform partners on hardware interface adaptation, scheduling of test time slots, and on-site operation arrangements.▎AI Technology Review: Are there any problems with multi-institution coordination? Can you give one or two specific examples?
Organizers:There are two particularly prominent difficulties:one is how to ensure evaluation results are reproducible across different computing environments；the other is how to coordinate robot equipment and personnel at different sites.For example, running the same evaluation code on GPUs of different architectures such as NVIDIA and AMD, if the underlying operator implementations and environment dependencies are not carefully aligned, results may show slight fluctuations. Therefore, it is necessary to adapt across multiple hardware environments, unify underlying configurations, and then conduct centralized review, to avoid misjudging platform differences as differences in model capability. Real-robot evaluation also involves venues, equipment status, and personnel arrangements in different regions, requiring task and data collection processes designed for different robot embodiments. Remote model services must also coordinate with on-site execution at millisecond-level timing; interface protocols, physical initial conditions, and execution logs all need to be clearly documented.▎AI Tech Review: There are many evaluation dimensions. Which can be automatically evaluated, and which rely on subjective scoring by experts? How do you ensure scoring consistency across institutions and tracks?
Organizers:The degree of automation differs across the three tracks, depending on whether the tests involve online interaction and on-site device operation.Track 1 mainly evaluates generative capability, with metrics computed mainly by automated processes, while manual review is retained for abnormal outputs and suspicious submissions.Track 2 examines the iterative performance of reinforcement learning policies in world model environments, and the evaluation process is also highly automated. Manual work mainly involves troubleshooting the connectivity of service calls and coordinating the interaction computation time slots of teams.Track 3 involves real-robot operation, requiring on-site personnel to participate in device reset, grasping observation, exception handling, and task success/failure determination, and to maintain a real-time connection with participating teams, with both parties communicating in real time to confirm test results, based on which success rates are tallied. Consistency is mainly ensured by adopting unified rules within the same track. Each team's input format, API specifications, computation budget, evaluation metric versions, and failure determination standards are all kept consistent. Different tracks examine different capabilities, and scores cannot be directly compared; what we focus onis ensuring that the evaluation conditions and determination standards within each track are consistent.。▎AI Tech Review: Are the leaderboards divided into a challenge leaderboard and a daily leaderboard? What is the update mechanism of the daily leaderboard? How often is it updated?
Organizers:The challenge leaderboard and the daily research leaderboard serve different purposes, and their scope of publication and update arrangements also differ.The challenge leaderboard is used to determine competition results and awards, with clear submission deadlines, leaderboard freeze rules, and award review procedures. To ensure competition fairness, some test data is not fully disclosed. The daily research leaderboard is open to academia in the long term; the code and evaluation protocols are fully open-sourced, and researchers can reuse them locally, providing a continuous performance reference for subsequent algorithm improvements.The challenge adds two new tracks, examining whether models can truly execute
▎AI Tech Review: WorldArena 2.0 adds Track 2 (MB-RL) and Track 3 (real-robot WAM) compared with 1.0. How did this idea come about? Why add exactly these two tracks? What were the underlying considerations?
Organizers:Adding these two tracks is meant to furtherexamine the role of models in interaction and real-world execution.1.0 mainly focused on embodied data engines and offline policy evaluation, still relying mainly on open-loop, static testing within simulated environments. Models were fine-tuned locally before testing, lacking real-time feedback and unable to self-correct during interaction. 2.0 aims to further examine two capabilities. One isclosed-loop interactioncapability: whether the policy can use the world model as an interactive environment, continuously learning and trial-and-error within the dynamic space generated by the model. The other isembodied generalizationcapability: whether skills learned in the model environment can be applied to the manipulation of physical robotic arms after Sim2Real transfer. The underlying consideration is that generating coherent videos does not mean the model can provide reliable physical feedback; running stably in simulation does not mean it will still execute properly on real robots. Therefore, it is necessary to make online interaction and real-robot deployment separate tracks.▎AI Tech Review: In a previous interview with us, Teacher Shang Yu frankly acknowledged the limitation of the leaderboard's "sim-to-real gap". Does Track 3, the real-robot track, directly respond to this problem? To what extent can it solve it?
Organizers:Real-robot evaluation incorporates actual sensing, contact conditions, and the device execution process into testing, providing a complement to simulation research. Factors such as friction changes, lighting interference, motor control latency, and rigid-body contact affect real-robot task outcomes. However, the tasks, environments, and robot embodiments currently covered by Track 3 are still limited, and will be continuously expanded and improved.▎AI Tech Review: Will there be further iterations to WorldArena 3.0? What dimensions of testing would be considered?
Organizers:In the process of advancing 2.0, we have already seen several directions worth continued research: higher-frequency, longer-durationClosed-loop interaction, after a misstep in execution, theSelf-correction, between robots of different configurationsGeneralization transfer, as well as in fine assemblyTactile and Multimodal Perception Fusion。▎AI Tech Review: Participation is selective — could it happen that strong players crowd into one track while another track is deserted, thereby weakening the overall representativeness of the rankings?
Organizer:This is related to technical requirements, R&D costs, software and hardware conditions, iteration time, and so on. Each track has its own distinctive features, reflecting the capabilities of different models at different levels.Ranking team profiles: which technical approach is better?
▎AI科技评论 (AI Tech Review): In all three tracks, the top two spots include both industry-backed and academic teams. What are the respective core strengths and weaknesses of each in world model innovation?
Organizers:Judging from this year's participating teams, universities and companies complement each other's strengths. University teams are more willing to trycutting-edge architectures and new objective functions, exploring with greater flexibility, while company teams have accumulated more experience inhigh-quality data reserves, large-scale compute scheduling, and system deployment, and theirengineering operationsare also more stable. This is our observation of this year's event and should not be generalized to all teams. The competition has also made the gaps each side needs to fill more concrete. Algorithmic innovation still needs to run stably in real engineering workflows; having data and compute advantages also requires accurate physical modeling and clear problem definitions before it can translate into results. Combining research with engineering capability helps shorten the path from proposing a method to putting it into practice.▎AI科技评论 (AI Tech Review): World models currently have four main technical routes overall, and this year's winning teams basically covered JEPA, spatial intelligence/structured 4D, and pixel-based generation routes, but teams on the JEPA route seem to have had the better odds: ShiQiWeiLai took first place in Track 1, and ChenHunXian ranked second in Track 2. As the organizers, how do you interpret these results?
Organizers:Technical routes in real systems are not entirely separate. Pixel-level autoregressive generation, representation learning in latent space, such as the JEPA idea, as well as spatial topology modeling and explicit physics engines, can be combined within a single system, and many leading teams adopt multi-module architectures. In addition, teams differ greatly in model parameter counts, pretraining data cleaning, action abstraction interfaces, and control loop frequencies. These results provide leads worth further research for embodied representation learning.▎AI科技评论 (AI Tech Review): Having witnessed so many leaderboard changes and the evolution of the competition, how developed do you think world models are now? What is the biggest bottleneck?
Organizers:Judging from this year's competition test results, in controlled, specific tasks, leading world models have initially developed the ability to respond to the causal effects of actions and can provide some help for downstream policy learning. But in longer-horizon complex interactions and contact-rich real robot manipulation, world models still face bottlenecks.Models cannot yet depict the state transitions caused after an action is applied accurately and consistently enough.Small errors in continuous prediction gradually accumulate and amplify; on real hardware, one must also handle millisecond-level action latency and the nonlinear force feedback caused by rigid-body contact.Continuously Advancing Research on Evaluation Methods
▎AI科技评论 (AI Tech Review): In our last interview, Professor Shang Yu mentioned the need to 'avoid WorldArena becoming a new path dependency for models' (everyone chasing high scores on WorldArena by catering to its metrics and tasks). Over two editions of the challenge and ongoing leaderboard updates, have you actually observed such a dependency? What are its specific manifestations?
Organizers:Optimizing against public rules is a phenomenon that is hard to completely avoid in benchmark operations. What we care about is whether score improvements truly reflect improvements in a model's general physical capabilities, or merely better fitting of a particular metric or a particular class of tasks. To reduce the latter's influence, we have mainly done three things:continuously tracking the correlation coefficient between scores and success or failure in real physical tasks, calibrating metrics that give high scores but do not reflect actual physical capability;continuously adding scenarios and tasks, expanding the OOD test distribution, and including more complex contact tasks;retaining an independent, closed blind test set, with key test segments kept strictly confidential.▎AI科技评论 (AI Tech Review): If you were to set one goal for WorldArena's next stage, what problem would you want it to focus on solving, and what role would you want it to play?
Organizers:In the next stage, we hope to more clearly assess a model's applicability in the real world: when researchers get a world model, they should be able to know in which physical scenarios it is reliable, and under what conditions and in which situations it tends to fail. The longer-term goal is,to let WorldArena help everyone improve their methods, so that world model capabilities generalize better to real-world scenarios。To make it convenient for everyone to gather, exchange ideas, and share conference news, we have specially created an IROS 2026 Conference Exchange Group! Join the group and you'll unlock these "perks"?
Conference Info Station: submission DDLs, formatting guidelines, review progress... key milestones delivered the moment they happen, so you'll never miss a deadline again and your submission preparation will be rock solid.
Peer Tea Chat: Peers from all over the country are in the group, sharing insights, exchanging ideas, and building connections—on the research road, we walk together.
Frontier Supply Station: the latest research findings and hot topics synced in real time, helping you sort out your submission strategy in line with the conference theme and giving your paper ideas a full tank of fuel.
On-site support team: (exclusive opening during the conference):Don't panic when you arrive at the venue either —
Route navigation: Venue layout, how to get to the lecture hall — ask anytime in the group chat;
Topic Group Reading: Popular sessions and Keynote discussions — you can catch up even if you missed them;
Poster compilation: A curated roundup of on-site poster highlights, save them all in one click and miss nothing;
Meetup Plaza: Dinner gatherings, interview meetups, networking sessions... Whether you want to chat, collaborate, or meet face-to-face, just start it in the group.
? Group entrance portal: Scan the QR code to join the group, or add WeChat Luyoyo_2026 with the note: IROS + institution/school + name + research direction.
In research and technology, the information gap matters.
Come on, stay one step ahead together!
Hop on board and explore the highlights of top global AI conferences
Exclusive access to:
Expert presentation slides
Full conference reports
Interpretations of popular papers
Interviews with rising academic stars
Scan the QR code above
or click "Read Original" to follow the special section.
Leiiphone (official account: Leiiphone)
Leiiphone (Leiphone.com)
This is an original article by Leiphone. Unauthorized reproduction is prohibited. For details, seethe reproduction notice。