While machine learning has long been a cornerstone of robotics, the latest AI revolution is reshaping how robots are taught. Cloud-based simulations have made training faster and more efficient than ever.
Step inside the Soft Robotics Lab at ETH Zurich, and you find yourself in a space that is part children's nursery, part high-tech workshop and part cabinet of curiosities. The lab benches are strewn with foam blocks, stuffed animals - including a cuddly squid - and other colourful toys used to train robotic dexterity. Piled up on every surface are sensors, cables and measurement devices. Skeletal fingers, on show in display cases or attached to powerful robotic arms, seem to reach out to grab you from every corner. The lab is home to a team of 19 robotics engineers, computer scientists, chemists and biologists, all working under the keen eye of Robert Katzschmann, Professor of Robotics in the Department of Mechanical and Process Engineering at ETH Zurich.
When it comes to designing next-generation robots, Katzschmann likes to take his cue from animals and human anatomy. His most recent robotic hands have fingers that are no longer powered by motors embedded in the joints but rather actuated by artificial tendons that move through rolling joints. The goal is to build robots that are soft, supple and agile. By replacing metals, screws and motors with hybrid bodies made of soft and rigid materials, his team is developing machines that can perform a wide range of tasks and adapt seamlessly to new environments.
Robots in action
This text appeared in the 25/04 issue of the ETH magazine Globe .
Adaptive hands
Katzschmann harnesses the power of artificial intelligence - though he still prefers the more precise term machine learning, emphasising that we're a long way from anything resembling a truly living intelligence. "In the past, we solved robotics problems through simplification and the use of physical models and control engineering," he says. "Today, we rely primarily on machine learning." This data-driven approach has permeated nearly every domain of robotics, from generative design with 3D simulation and skills acquisition from video to algorithm-based motion control. "Around half of my group is actively engaged in applying and refining machine-learning methods," Katzschmann says.
Traditional methods, like control engineering, excel at structured, repetitive processes, such as those found on the factory floor. But they fall short when it comes to chaotic environments and unstructured tasks. As Katzschmann points out, something as simple as sorting different glass bottles into a crate remains a formidable challenge for robots because of the bottles' varied shapes and sizes. To tackle this, his group developed a robotic hand with 21 degrees of freedom. Trained through a combination of reinforcement and imitation learning, this dexterous hand has been made part of a larger system that, together with a robotic arm, now offers an impressive 28 degrees of freedom. To teach the robot, researchers wear a glove fitted with motion sensors and a camera. As they demonstrate how to grasp a bottle, their movements are recorded by external cameras. This rich dataset, sometimes augmented with virtual reality images, is used to train what is known as a transformer model - an architecture similar to the large language models that power modern AI. Once trained, the robotic hand can also pick up unfamiliar objects and move them to the right place. "With traditional methods, we would have had to create a 3D point-cloud model of the environment and then program every single finger position required to grasp the bottle," says Katzschmann. "Even the slightest change in position of the bottles or the crate would have left the robotic hand not knowing how to respond." But that's no longer the case. "The movement required to pick up a bottle is now entirely learned, and that makes the hand remarkably adaptive," he explains. In 2024, this research spawned Mimic Robotics, an ETH spin-off set up by Katzschmann and four of his former doctoral and Master's students. The fledgling company aims to revolutionise manufacturing and logistics with its AI-controlled robotic hands.
Learning in the cloud
Stelian Coros is a computer scientist who develops algorithms for robotics, visual computing and computer-aided manufacturing. His main focus is the software that forms the robot's brain, and his research over the past decade has been shaped by advances in deep learning, a form of machine learning that uses artificial neural networks. "We've reached the stage where we have enough data and computing power to use deep learning for specific applications in robotics, such as automatic object recognition in images," he says.
Neural networks also form the basis of another type of machine learning known as reinforcement learning, a technique in which robots improve through trial and error. Researchers reward a robot for desired outcomes - such as moving forward without falling - and the robot continuously refines its actions to maximise its score. "It's kind of learning by doing, much like people pick up how to play tennis," says Coros. "It's not enough for robots to watch YouTube videos of people performing a task. They have to try it out for themselves." His team therefore generates vast amounts of training data through teleoperation. This is what enables a robot to replicate a human operator's movements. They also use motion-capture technology from the animation industry to record human actions. Harnessed by suitable algorithms, this data enables robots to perform context-appropriate, human-like movements - a crucial prerequisite, Coros argues, for seamless human-robot interaction.
Parallel training
At the Robotics Systems Lab (RSL), led by Professor Marco Hutter, researchers also rely on reinforcement learning, but applied on a large scale in virtual settings. "We use simulations to train thousands of robots at the same time," says Cesar Cadena, a senior scientist in the lab. "In one hour, we now generate as much data as we used to produce in a year." These simulations have been made possible by huge advances in microchips and graphics processors. Parallel processors can execute thousands of tasks simultaneously and are fundamental to AI applications, which is why RSL also works closely with NVIDIA, one of the world's biggest developers of graphics processors and chipsets. This has already given rise to two theses directly tied to the California-based company.
Virtual reinforcement learning takes place in the cloud and requires immense computing power. If in continual learning mode, however, this dependency could compromise a robot's autonomy. A factory robot, for example, can stay connected to the cloud to optimise its performance of complex tasks. But what happens to a rescue robot searching for survivors in a remote disaster zone? If there is no network coverage, how can it make rapid decisions? To circumvent this problem, researchers install some of the computing capacity in the robot itself, along with pre-generated data from the cloud. "We sacrifice some processing power," Cadena says. "But for clearly defined tasks, it's usually still good enough."
The goal: versatile robots
Does the current boom in AI herald a robotics revolution - or more of an evolution? Coros believes it's the latter. "The data needed for AI and the data needed for robotics are fundamentally different in kind," he says. A robot has a body and must learn through physical interaction to generalise movements and apply them in a variety of environments. By contrast, AI achieves generalisation through an unending stream of data - primarily text, but also images and videos. Some in the robotics community continue to pursue a purely data-driven approach, training robots with terabytes of human motion data. "But that simply isn't viable," Coros insists. He points to research groups that developed robots to fold shirts, noting that it took some 10,000 hours of demonstration data before a robot was able to complete the task - and even then it made mistakes:"If it requires that much data to learn a single skill, then it's fundamentally unscalable as an approach."
His group has therefore taken a different path, combining learned data with physical models to fill gaps in demonstrations. Coros cites the example of a robot arm throwing a ball: "We understand the physics of how a ball moves through the air." A robot can use these physical laws to adjust its throw so that the ball lands precisely where intended. "And we can do that without requiring masses of data," says Coros. In 2023, Coros partnered with former doctoral students to launch the spin-off Flink Robotics. The company uses AI-powered image processing and physical models to make standard industrial robot arms more intelligent, enabling them to package, unload and sort materials with greater precision. Swiss Post, Flink Robotics' first client, plans to use this technology to automate its parcel operations.
Tendons trump motors
Back in the Soft Robotics Lab, biologists are developing cell tissue for artificial tendons, while chemists inject life into artificial muscles by means of electrical impulses. Katzschmann is convinced that traditional, motor-driven robots are reaching their limits when it comes to generalisation, no matter how sophisticated their AI. "Those kinds of systems simply won't be adaptable enough to deal with all the situations they'll encounter in the real world." For him, the body is as important as the brain, which is why he is working on musculoskeletal robots that mirror natural designs. "Muscles provide softness," he says, "and the skeleton provides the load-bearing capacity needed for complex physical work." Nature has engineered remarkably stable and versatile systems without the need for motors or metals. "That should be our model," he argues.
