Cobots with the muscle of imagination

From October 19 to 25, Hangzhou, the economically powerful metropolitan region in eastern China, was the hotspot of the robotics world. More than 5,000 researchers and industry leaders gathered at IROS 2025 – the International Conference on Intelligent Robots and Systems – and made it clear that our ideas about robot colleagues are always just a snapshot in time.

“Staging the Machine: Not Built for Work, Built for Wonder” was the keynote address by Dennis Hong, professor at the University of California, and it showed, like many other contributions, that we will continue to be amazed in the future.

One of the highlights in Hangzhou was RoboSense's Robot Manipulation Eye (AC2), a camera that can hardly be described as a sensor anymore, but rather as a sixth sense. Its direct time-of-flight technology is said to deliver millimeter-precise depth images, even in bright light and when in motion. The developers of the Robot Manipulation Eye say that it enables “perception like a living being” – and that's not just marketing talk. In the future, the Robot Manipulation Eye will enable gripper arms, autonomous vehicles, and service robots to not only see their surroundings, but to truly understand them.

Artificial intelligence with tactile sensitivity

The topic of “understanding” was featured in numerous workshops at IROS 2025. Robots no longer learn movements alone, but entire concepts – with multimodal learning systems and “human-in-the-loop” training – and combine visual, linguistic, and haptic signals to accomplish tasks in industrial environments as well as in real life. Opening a stubborn medicine package or sorting fruit according to ripeness is therefore no longer a unique human skill. And the potential associated with this seemingly unspectacular insight was the top topic in Hangzhou.

Robots without an elbow mentality

One of the fascinating developments presented at IROS 2025 is social navigation. A team from Xiaomi demonstrated that a service robot that simply avoids people in confined spaces for safety reasons is far less empathetic, even polite, than a model that tries to understand people's intentions. In practice, this means that the robot pauses briefly and makes a small detour to maintain distance without appearing dismissive. The system is based on RGB-D sensors and a module for “proactive risk assessment,” as the developers call it. So the future of robot interaction may be more than just efficient; it could also be empathetic – or at least appear to be.

Research and practice are converging

A particularly stable trend at IROS 2025 was that robots are increasingly leaving their sterile test halls and taking on tasks assigned to them by humans in agriculture, logistics, disaster control, and, last but not least, medicine – and this is happening with steadily increasing confidence in the competence of AI-controlled colleagues, so that smart machines are pushing their way into real environments everywhere. Particularly impressive here are the medical “continuum robots,” whose bodies consist not of rigid joints but of flexible, bendable structures, with their surgical capabilities.

Robots with self-healing powers

Since such “soft” robots inspired by biological systems not only perform well in the operating room but are also more susceptible to damage due to their structure, Bram Vanderborght from Vrije Universiteit Brussel presented robots with self-healing capabilities at IROS 2025. “Our technological advances enable these robots to repair themselves, improving their durability and significantly extending their service life. This innovation not only increases reusability, but also allows for recycling processes and is based on bio-based materials – an important contribution to sustainability,” Vanderborght noted in his keynote speech. “Our groundbreaking innovations include self-healing gripper arms with integrated sensors that can not only detect damage but also actively respond to it. We are currently further developing the technology to found a deep tech spin-off called Valence Technologies – with the aim of bringing self-sealing suction cups and self-healing bicycle and car tires to market,” Vanderborght continued.

From semantic understanding to safe autonomy

German computer scientist and university lecturer Angela P. Schoellig, Alexander von Humboldt Professor of Robotics and Artificial Intelligence at the Technical University of Munich, stated in her keynote speech in Hangzhou that foundation models and large-scale learning are fundamentally changing robotics. Today, robots can be programmed using language, visual impressions, and demonstrations – instead of laboriously handwritten code. But according to Schoellig, this new paradigm raises a key question: How can safety be ensured in open, dynamic environments? To answer this, Schoellig presented current work that combines semantic understanding and generative control strategies with safety-critical control frameworks. Examples included safe navigation in crowds using diffusion models and human motion prediction – similar to the Xiaomi service robots mentioned above – semantic safety filtering for manipulation tasks, and swarm control via natural language instructions with built-in protection mechanisms. “These results highlight both the new opportunities and challenges associated with the use of AI on an internet scale in robotics,” said Schoellig. She concluded by presenting a roadmap for trustworthy, scalable, and secure autonomous systems in human-centered environments.