On June 5, the five-day IEEE International Conference on robotics and automation (ICRA 2026) officially concluded in Vienna, Austria. As one of the most inFluential global gatherings in the Robotics SECtor, this year's conference attrACTed over 8,000 attendees on-site. Out of 4,947 valid submissions, 1,882 papers were accepted, resulting in an acceptance rate of 38.04%. Compared to 2023, which saw 3,125 submissions and a 43% acceptance rate (with 1,345 accepted papers), this year's conference experienced a massive surge in submissions alongside a nearly 5-percentage-point drop in the acceptance rate. This reflects a broader industry shift: as the embodied intelligence sector becomes increASIngly vibrant, the threshold for academic publication has been rAIsed.
This trend also mirrors the changing macro-environment of the industry. Over the past 12 months, the world has witnessed a startup boom in Embodied Intelligence, with talent accelerating from smart vehicles, AI, and consumer electronics. Meanwhile, Unitree's push for an IPO and OpenAI's return to robotics have driven capital and industrial enthusiasm to unprecedented heights. At this critical juncture, ICRA cARRies significance far beyond academic exchange; it serves as a collective roll call for the dawn of the Physical AI era.
Best Papers Unveiled: Chinese scholars Shine
In the highly anticipated Best Paper Award selection, Assistant Professor Shiguanya's team from Carnegie Mellon University (CMU) won both the Best Paper Award and the Best Manipulation and Motion Paper Award for their paper, OmniRetarget: Interaction-Preserving Data Generation for Humanoid Whole-Body Loco-Manipulation and Scene Interaction. This rare "double win" in ICRA history highlights a groundbreaking APProach. Traditionally, training humanoid robots relies on mimicking human movements, but structural differences often lead to foot slipping, CLIpping, and lost interaction details. The team introduced OmniRetarget, a motion retargeting tool that preserves all contact and spatial interactions to generate compliant motion trajectories. It can rapidly expand training data for various robots, terrains, and objects based on a single motion segment. In real-world tests, generating over eight hours of high-quality motion Samples allowed the Unitree G1 robot to independently complete a 30-second continuous parkour and handheld manipulation task using a simple reward mechanism.
The Best Automation Paper Award went to Professor Hu Ruizhen’s team from Shenzhen UniveRSIty for IMR-LLM: Industrial Multi-Robot Task Planning and Program Generation using large language models. Addressing the complex logical constraints of industrial multi-robot collaborative Operations that ordinary LLMs struggle with, the team proposed the IMR-LLM scheduling Framework. It uses LLMs to build task logic graphs for optimal overall planning and generates executable robot programs via process trees. Supported by a hierarchical industrial robot testing dataset, the system's comprehensive performance significantly surpasses existing solutions.
Professor Quan Quan’s team from Beihang University won the Best Field and Service Robotics Paper Award for Planar-Sector LOS Guidance for Interception of agile Targets with Robotics Lifting-Wing Quadcopters. Tackling the challenge of stably tracking and intercepting high-speed, unpredictable targets in search, rescue, and security scenarios, they proposed a novel "Planar-Sector Line-of-Sight (LOS)" guidance method. This allows lifting-wing quadcopters to dynamically adjust flight paths based on target movements, enhancing interception stability while maintaining high-speed maneuverability. Experiments confirmed its ability to achieve precise tracking and long-range interception of highly AGIle targets in complex environments.
The Best Student Paper Award was awarded to Xiao Chenxi’s team from ShanghaiTech University for ETac: A Lightweight and Efficient Tactile Simulation Framework for Learning dexterous manipulation. Addressing the trade-off between simulation fidelity and speed in flexible tactile sensing, the team developed ETac. This lightweight model simulates soft-body deformation with accuracy approaching Professional finite element analysis while reproducing real tactile sensor data. A single RTX 4090 GPU can run 4,096 simulation scenes simultaneously at 869 FPS. Training a grasping model using only tactile data (without vision) achieved an 84.45% success rate across four object categories, enabling low-cost, large-scale training of robotic tactile manipulation Skills.
Data Remains a Core Challenge, But It Is Not a Panacea
Entering 2026, discussions around data in the embodied intelligence industry have intensified. While LLMs succeeded due to massive datasets, robotics remains bottlenecked by data scarcity, making data collection a primary focus. At this ICRA, UC Berkeley Professor Ken Goldberg delivered a keynote emphasizing data collection and data flywheels.
Regarding the highly anticipated "ChatGPT moment" for robotics, Goldberg expressed General agreement that data will solve robotic challenges just as it did for LLMs, but the key lies in when. He presented a stark contrast: at an aveRAGe human reading speed, it would take 100,000 years to read all the data used to train LLMs, whereas the robotics field currently has only a few years' worth of scale. This massive gap leads many to believe that sufficient data will effortlessly solve robotic model issues. However, Goldberg argues that data alone is insufficient; traditional engineering remains crucial. Extensive engineering design, modular systems, and meticulous debugging, combined with data, are the keys to making systems truly operational.
Furthermore, compared to synthetic simulation data or teleoperation, Goldberg emphasized data generated in real production environments—deploying robots and collecting working data during deployment. This aligns with the current global focus on the "data flywheel": mass deployment generates real-world data, which feeds back to optimize models, and better models subsequently promote high-quality robotic deployment.
Three Major Trends
After a year of industry development, a clear shift emerged at this year's ICRA: embodied Intelligence is transitioning from single-point capabilities to full-stack, scalable capabilities focused on real-world deployment. In this process, physical AI has become the unifying narrative. Physical AI emphasizes not just completing specific tasks, but teaching AI to underStand the real physical world. Over the past year, the focus has shifted toward how robotic task capabilities can generalize to the physical world. Consequently, keywords like world models, VLA (Vision-Language-Action), Sim-to-Real, and real deployment data appeared frequently. In a sense, robotics is experiencing its own large model moment, accelerating the creation of Physical AI systems capable of continuous learning and generalization.
1. Scalable Data Expansion Becomes Core
Data has never received as much attention at ICRA as it did this year. Amid the consensus of severe data shortages, the industry is exploring how to produce real-world data at scale and low cost. As the pursuit of generalization accelerates, the data gap widens. The industry is simultaneously betting on multiple data pathways, including teleoperation, real-machine deployment, human data, and simulation. More companies are establishing "data factories" to collect real operational data and deploying robots in factories, logistics, and retail to form data flywheels. Meanwhile, in simulation data generation, from the Best Paper OmniRetarget to large-scale tactile simulation for Dexterous Manipulation, researchers are trying to solve the same problem: expanding limited real data into sufficient training scale.
Data has never received as much attention at ICRA as it did this year. Amid the consensus of severe data shortages, the industry is exploring how to produce real-world data at scale and low cost. As the pursuit of generalization accelerates, the data gap widens. The industry is simultaneously betting on multiple data pathways, including teleoperation, real-machine deployment, human data, and simulation. More companies are establishing "data factories" to collect real operational data and deploying robots in factories, logistics, and retail to form data flywheels. Meanwhile, in simulation data generation, from the Best Paper OmniRetarget to large-scale tactile simulation for Dexterous Manipulation, researchers are trying to solve the same problem: expanding limited real data into sufficient training scale.
2. Fierce competition in Dexterous Manipulation
This year, dexterous manipulation has become one of the most fiercely contested battlegrounds in robotics. For robots, walking is merely mobility; true Productivity lies in their hands. Whether in factory assembly, home organization, logistics sorting, or industrial maintenance, tasks essentially rely on grasping, twisting, pressing, inserting, and moving. These seemingly simple actions are incredibly difficult for robots, requiring not only visual understanding but also complex technologies like force control, tactile feedback, and contact modeling. The Best Student Paper (ETac) on efficient tactile simulation, along with numerous workshops on dexterous hands and dual-Arm collaboration, dedicated significant time to researching how robots can truly gain a "sense of touch." Compared to simple grasping and carrying, the industry now emphasizes complex, fine manipulation tasks. These capabilities are unavoidable for real-world deployment and are likely to be the most competitive and critical mainline in embodied intelligence over the next two years.
This year, dexterous manipulation has become one of the most fiercely contested battlegrounds in robotics. For robots, walking is merely mobility; true Productivity lies in their hands. Whether in factory assembly, home organization, logistics sorting, or industrial maintenance, tasks essentially rely on grasping, twisting, pressing, inserting, and moving. These seemingly simple actions are incredibly difficult for robots, requiring not only visual understanding but also complex technologies like force control, tactile feedback, and contact modeling. The Best Student Paper (ETac) on efficient tactile simulation, along with numerous workshops on dexterous hands and dual-Arm collaboration, dedicated significant time to researching how robots can truly gain a "sense of touch." Compared to simple grasping and carrying, the industry now emphasizes complex, fine manipulation tasks. These capabilities are unavoidable for real-world deployment and are likely to be the most competitive and critical mainline in embodied intelligence over the next two years.
Conclusion
If the development of the robotics industry over the past few years is viewed as a marathon, ICRA 2026 represents a highly special milestone. Following a rapid inflUX of capital, startups, and industrial resources, there is now a relatively clear consensus on the problems the embodied intelligence sector must solve in the future. The robotics industry has become more pragmatic, focusing on specific issues: how models generalize to unfamiliar environments, how to acquire data at scale, and how to improve the success rate of dexterous manipulation. Many of these problems are not glamorous and can even seem trivial. However, this is precisely the signal that the industry is accelerating forward. Once in the real world, robots face a series of specific, complex, and unavoidable challenges. This year's ICRA is increasingly centered around these seemingly fragmented but vital issues.
Comments & Questions (0)
No comments yet
Be the first to comment!