Author | Ma Xiaoning
Editor | Cen Feng
On September 30, local time, the main IROS 2026 conference concluded in Pittsburgh, USA, and the major awards were announced on the last day. According to AI technology reviews on site, this year’s Best Paper Award was awarded to a study on robots’ “long-term memory.” Yumin Lee, Hyoseok Ju, and Giseop Kim won the best paper for their paper titled “LT-Mem: Volatility-Aware Spatio-Temporal Memory for Lifelong Scene Understanding.” The Best Student Paper Award went to Pei-An Hsieh, Fengjun Yang, Nikolai Matni, and M. Ani Hsieh wins. Their research focuses on how multiple quadcopter drones handle complex airflow interference when flying in close formation. Two other impressive awards go to humanoid robots. One project has the humanoid robot learn real tennis movements, while another has it hold a tray while walking on two legs, keeping the objects on the tray stable. If we compare this year’s award-winning papers with those in the Best Paper Finalist category, areas such as long-term memory, VLA, world models, 4D scene understanding, visual-tactile operations, flight control, humanoid robots, soft robots, and bio-molding robots are all included in the list. Instead of simply discussing what new actions robots have learned, this year’s award-winning works begin to address a more practical issue. Once robots enter real-world environments, how can these capabilities function continuously and reliably?
01
The best paper gave robots “long-term memory”
What should a robot remember if it lives in the same place for a year? This is a core question behind the best paper at IROS this year. The award-winning paper by Yumin Lee, Hyoseok Ju, and Giseop Kim is titled “LT-Mem: Volatility-Aware Spatio-Temporal Memory for Lifelong Scene Understanding”. For most robot systems today, “understanding a scene” mainly refers to understanding the scene at this very moment. The robot entered the office and saw the table, chairs, and printer—enabling it to create a scene representation. The next day, the chairs were moved away, and it could update its map based on new observations. This led to problems. If new states continuously overwrite old ones, then although the robot knows “what it is like now,” it gradually loses “what it was like in the past.” LT-Mem refers to this phenomenon as temporal amnesia, or time-based forgetting. For robots that operate over long periods, history can be extremely important. A wall hardly moves, a office chair is occasionally moved, people’s backpacks may appear in different locations every day, and a parking space may follow a certain periodic pattern of occupancy. The robot not only needs to recognize these things, but also needs to know what changes, how often they change, and how they changed in the past. Therefore, LT-Mem does not simply continue to expand the model, but designs a layered long-term memory structure, which includes three levels: Live Memory, Delta Memory, and Meta Memory. Live Memory saves the state of the current world, Delta Memory records the changes that have occurred in the scenario, and Meta Memory further summarizes patterns from long-term history. In this system, a particularly important concept is Volatility, or variability. The robot determines whether a particular object is a relatively stable component of the environment or a subject that undergoes frequent changes, and based on this, it decides how to organize its memory. So, the problems that robots face are no longer just “where is the green chair?” They can also evolve further to “where did this green chair appear in the past?” “When is it most likely to become empty?” “Is the position of this object stable enough to serve as a reference for long-term navigation?” This is also an important difference between lifelong scene understanding and ordinary scene understanding. What robots face is no longer a static picture, but a world with history. Finally, LT-Mem emerged from the 10 Best Paper Finalists this year. To some extent, this result also reflects a more and more realistic issue that robotics research is facing. If robots are actually expected to work continuously for months or even years in the future, how should they understand time?02
From "flying close by", to playing tennis and carrying trays
If the Best Paper discusses time measured in months or even years, then the Best Student Paper of this year examines a problem that occurs within milliseconds. “Flatness-Preserving Residual Learning for Real-Time Tight Quadrotor Formation Flight” by Pei-An Hsieh, Fengjun Yang, Nikolai Matni, and M. Ani Hsieh received the IROS 2026 Best Student Paper Award. Flying a formation of multiple drones is not a new issue; the real challenge lies in what to do when they fly very close together. The downwash airflow generated by the quadcopter affects nearby aircraft. As the distance between drones becomes even smaller, these complex aerodynamic disturbances become quickly unacceptable to ignore. Therefore, researchers have introduced physics-informed residual dynamics learning, allowing the learning model to compensate for the aerodynamic effects that traditional dynamic models struggle to accurately describe, while preserving the system’s differential flatness, thus continuing to utilize efficient trajectory planning and control methods. This paper also made it to the Best Paper Finalist list and ultimately won the Best Student Paper Award. However, the robots that are most easily understood by ordinary readers on the award stage this year are probably two other ones, as one is playing tennis and the other is carrying a tray. Zhang Zhikai, Haofei Lu, Yunrui Lian, Ziqing Chen, Yun Liu, Chenghuai Lin, Han Xue, Zeng Zi Cheng, Zekun Qi, Shaolin Zheng, Qing Luan, Jingbo Wang, Junliang Xing, Wang He, Li Yi, etc.’s “Learning Athletic Humanoid Tennis Skills from Imperfect Human Motion Data” won the Best Entertainment and Amusement Paper Award. Tennis is a rather demanding full-body motion task for humanoid robots. It is not just about the robotic arm swinging the racket; the robot also needs to observe the ball, move its body, adjust its center of gravity, control the timing of the strike, and at the same time coordinate its legs, torso, and arms. One key issue in this work is already mentioned in the title: Imperfect Human Motion Data. In reality, human motion data is never complete, standard, or precise. If each time one sports skill is learned requires the collection of a large amount of high-quality motion capture data, then the data itself becomes an important bottleneck in enhancing the skills of humanoid robots. This work attempts to enable robots to extract and combine motor skills from imperfect human motion data, and ultimately implement tennis tasks on humanoid robots. Another award-winning paper is not as impressive; the robot simply walks with a tray in hand. Anlun Huang, Zhenyu Wu, Soofiyan Atar, Yuheng Zhi, and Michael C. Yip’s “SteadyTray: Learning Object Balancing Tasks in Humanoid Tray Transport Via Residual Reinforcement Learning” received the Best Paper Award in Mobile Manipulation. But “walking” and “walking with a load” are actually completely different challenges. When bipedal robots walk, body movements and impacts are inevitable. If the objects on the tray are not secured, these disturbances will be directly transmitted to the objects. The fact that the robot itself does not fall does not mean the cup won’t fall. SteadyTray’s approach is not to retrain a large model that handles all problems, but to separate locomotion from payload stabilization. The existing lower limb strategy is responsible for walking, while the additional residual policy learns how to use upper body movements to counteract the disturbances caused by the tray. What it represents is also a significant change in humanoid robots in recent years. Research is moving from mere locomotion towards loco-manipulation: robots not only need to move their bodies, but also must complete tasks during the movement.03
Navigation, dressing, farming, underwater operations
The robot enters a more specific world
If long-term memory, humanoid robots, and drones represent the most prominent directions in this year’s award results, then the other special awards demonstrate how diverse real-world environments robots are entering. Yi Du, Taimeng Fu, Zhipeng Zhao, Shaoshu Su, Zitong Zhan, Qiwei Du, Zhuoqun Chen, Bowen Li, and Chen Wang’s “VL-Nav: A Neuro-Symbolic Approach for Reasoning-Based Vision-Language Navigation” won the Best Paper Award in Cognitive Robotics. This work points toward the highly important Vision-Language Navigation in recent years. Robot navigation is shifting from traditional “moving from point A to point B” to understanding environments, objects, and spatial relationships based on linguistic descriptions, and then taking action accordingly. VL-Nav chooses a neuro-symbolic approach, attempting to combine the powerful perception capabilities of neural networks with structured symbolic reasoning. Another award-winning paper even redesigned something we use every day, namely a zipper. “Design, Modeling, and Performance of a Robotic Zipper for Self-Donning Clothing” by Amanda Weckerly, Wooseok Kim, Megan Coram, Ellen Li, Sophie Lin, Gabriel Unger, Isabella Szabo, Allison M. Okamura, and Cynthia Sung received the Best Paper Award in Robot Mechanisms and Design. What’s interesting is that the researchers didn’t just make the robot more complex; instead, they redesigned the physical world itself to make clothing easier to wear automatically. This is also a very different aspect of robotics and pure AI research. Solving a problem does not necessarily require a larger model; sometimes, better mechanical design is the answer. In the field of agricultural robots, the paper “SSL-Semantic-NBV: A Self-Supervised Learning-Based NBV for Target-Aware 3D Reconstruction in Agricultural Robotics” by Jianchao Ci, Katarina Smolenova, Xin Wang, Axel Streit, Eldert J. van Henten, and Gert Kootstra received the Best Paper Award on Agri-Robotics. This work focuses on the Next-Best-View problem in three-dimensional reconstruction in agricultural scenarios. Robots do not continue filming aimlessly; instead, they need to determine where to move next in order to obtain the most valuable new information about the target. In the field of underwater robots, Michele Grimaldi, Yosaku Maeda, Hitoshi Kakami, Ignacio Carlucho, Yvan R. Petillot, and Tomoya Inoue’s “Contact-Aided Factor-Graph Localization for Underwater Sampling” received the Best Paper Award in Safety, Security, and Rescue Robotics. In an underwater environment where both GPS and stable visual information are limited, researchers attempt to use the contact between robots and the environment to assist in positioning. For real robots, “what they touch” can also serve as perceptual information. Additionally, “Acoustic-Aware Texture Optimization for Ultrasound-Based Capsule Pose Estimation” by Yizhao Qian, Yuzhuo Wang, Chenxi Liu, Jiayuan Luo, Wenwei Gao, Yuan Yixuan, Meng Qinghu, and Li Liu received the Best Application Paper Award. In the field of industrial robotics, the paper “Picking Bins Empty: A Hierarchical Hybrid Approach with Online Self-Learning of Grasp Points for Reliable Industrial Bin-Picking” by Florian Töper, Samarth Kishor Yelvande, Jan Niklas Ewertz, Rudolph Triebel, and Peter Ohlhausen received the Best Paper Award for Industrial Robotics Research for Applications. Its title is particularly direct: “Picking Bins Empty”, which means picking empty boxes. What industrial robots really need to prove is not just “I grabbed it once successfully,” but whether they can continue to perform the grabbing task despite obstructions, stacking, posture changes, and new states that arise from repeated operations, until the box is truly emptied.04
10 top paper candidates,
What is being rewarded in robot research?
Only LT-Mem won the Best Paper award, but the 10 papers that advanced to the final stage may better indicate what IROS 2026 is focusing on. In addition to LT-Mem and the quadcopter formation work that won the Best Student Paper, there is also the Shallow-π research by Boseong Jeon, Yunho Choi, and Taehan Kim: knowledge distillation based on Flow-based VLA. “VTAP Gripper: Synergizing Fingertip Sensing and a Visuo-Tactile Active Palm for Dexterous In-Hand Manipulation” by Yuhao Zhou, Sheeraz Athar, Hu Zhixian, Huang Binghao, Li Yunzhu, Juan Wachs, and Yu She combines fingertip sensing with a visuo-tactile active palm, aiming at more complex delicate hand operations. “Streaming Gaussian Encoding for 4D Panoptic Occupancy Tracking” by Maximilian Luz, Thomas Nürnberg, Yakov Miron, and Abhinav Valada advances Gaussian representation to dynamic 4D scene understanding. “DAWN: Noise-Robust Quadruped Parkour Via Depth-Denoising World Models” by Yohan Choi, Min-Jun Kim, Jin-Sung Kim, Yong-Jae Kim, and Youn-Hee Han applies the Depth-Denoising World Model to quadruped robot parkour. Several other candidate papers involve three-dimensional reconstruction of glass surfaces, multi-modal fusion of EEG-EMG signals, magnetic soft robotics, and implanting biological manufacturing tissues with robot assistance. For example, the paper by Jiamin Zheng, Jingwen Yu, Guangcheng Chen, and Hong Zhang, “Enhancing Glass Surface Reconstruction Via Depth Prior for Robot Navigation”, directly addresses the challenging issue of glass surfaces in robot vision. “Autonomous Multimodal Locomotion Control of a Magnetic Strip Soft Robot” by Bin Wang, Shengming Luo, Jiansheng Du, Yuanbiao Ma, Xuanyu An, and Qianqian Wang has shifted the research focus to magnetic soft robots with completely different structures and movement mechanisms. SeungTaek Hong, Jaewon Byun, Daekeun Kim, Jinah Jang, and Keehoon Kim’s “Toward Autonomous Biofabricated Tissue Implantation: A Non-Contact Robotic Manipulation System for Soft and Fragile Modules” further brings robots into the scenario of operating on soft and fragile biofabricated tissues. What makes this list interesting is that there is no single technology approach that dominates all robotic challenges. VLA, World Model, Memory, Gaussian, Visuo-Tactile, and Residual Learning appear together, and a large amount of work is still being done to seriously address aerodynamics, contact, mechanical structure, soft material, and real-time control. In the field of robotics, a larger model cannot automatically solve all problems.05
When robots truly enter the real world
In addition to the paper award, IROS 2026 also announced a series of individual awards. Wang Zhidong from Chiba Institute of Technology in Japan received the IROS Distinguished Service Award, and the award was given for his long-term contributions to organizing the IROS conference. Roberto Martin-Martin from UT Austin and Roberto Calandra from TU Dresden received the Toshio Fukuda Young Professional Award. The former received the award for its focus on learning-based manipulation, contact-rich control, and physically interactive robot systems; the latter won due to its contributions in sample-efficient robot learning and the introduction of advanced tactile perception into robot intelligence. Ralph Hollis was awarded the Emeritus Keynote Award for Research Achievement. The conference specifically mentioned his long-standing contributions to robot research and his role in training several generations of robotics researchers. In addition, Alper Yegenoglu and Ariyan Bighashdel received the Outstanding Associate Editor Award, while Dingsheng Luo and Long Zeng received the Outstanding Reviewer Award. When considering these individual awards and paper awards together, the award winners list for IROS this year is not a concentrated victory in a particular model or direction. LT-Mem begins processing the robot’s long-term memory, DAWN applies the world model to quadruped robot movement, VTAP Gripper promotes visual and tactile integration, and humanoid robots start tackling complex tasks such as tennis and carrying trays. Meanwhile, robots are entering agricultural fields, factories, and underwater environments, and also face more complex real objects such as glass, clothing, and soft tissues. The technical approaches used in these works are different, but they all point toward a common trend: robot research is moving from “gaining a capability” to “continuously using that capability in the real world”. If we look for a common thread in the award results of IROS 2026, it may not be simply “embodied intelligence continues to develop”. When robots truly move into the real world, what they need is not just a larger model, but also the ability to remember the past, understand change, perceive contact, and maintain stable control over their bodies under real-world dynamic constraints. From walking, to playing tennis, to holding a tray, and eventually being able to long-term understand a constantly changing world, robots are moving from simply demonstrating a capability to becoming a more complete system.To facilitate group communication and sharing of meeting updates, we have specifically created the IROS2026 participation and communication group! What benefits will you unlock by joining the group?
Meeting Information Station: Submission DDLs, format guidelines, review progress… key updates delivered immediately, so you won’t miss the deadline anymore. Your submission preparation is well under control.
Peer Discussion Meeting: All peers from all over are in the group. We share insights, exchange ideas, and build connections—we travel together on the path of scientific research.
Frontier Supply Station: Latest research findings and hot topics are synchronized in real time. Combined with the conference theme, it helps you clarify the flow of your submission and provides inspiration for your paper.
On-site support group: (activated only during the meeting):You don’t need to panic when you arrive at the venue—
Route Navigation: How to reach the venues and the lecture hall, ask in the group at any time;
Topic Discussion: Popular sessions and Keynotes are available for viewing even if you missed them;
Poster compilation: A summary of key on-site posters, save it with one click without missing anything;
Yueju Square: Meal gatherings, interview sessions, communication forums… If you want to chat, collaborate, or meet in person, it can be started right here in the group.
? Group entry link: Scan the code to join the group, or add WeChat Luyoyo_2026, and note: ECCV + institution/school + name + research direction.
In scientific research or technology work, information gap is very important.
Come on, take a step ahead!
Get on the bus, and I’ll show you the highlights of global AI conferences
Exclusive access:
Expert Speech PPT
Full Report of the Conference
Interpretation of Popular Papers
Interview with Academic Rising Stars
Scan the QR code above
Or click “Read the original article” to follow the Leifeng Net (public account: Leifeng Net) section.
Leifeng Net original article, reproduction is prohibited without authorization. For details, see Reproduction Guidelines.