Robots have long been sorted by their shape. Articulated arms, wheeled rovers, humanoid frames—these physical characteristics have served as the primary way we organize and understand robotic systems. But a new approach proposes something radically different: classifying robots not by how they look, but by how well they perform the tasks we need them to do.

Researchers have introduced the "tree of robots," a taxonomy that connects robot design to real world performance. Inspired by Darwin's tree of life, this classification system evaluates robots based on their fitness for specific processes, particularly those involving physical interaction with objects and environments. The goal is to bridge the gap between a robot's morphology and its actual capability to execute industrial tasks, from delicate assembly work to forceful grinding.

This new framework offers a systematic way to measure and compare robot capabilities across different platforms. The team tested 11 commercial industrial manipulators using a carefully designed set of metrics, revealing dramatic differences in performance that existing classifications overlook. The findings are now published as an open access tool, inviting researchers worldwide to contribute measurements and expand the database.

From Shape to Function

Traditional robot classification has focused on kinematic structure—the number of joints, the arrangement of links, whether the base is fixed or mobile. While useful for understanding mechanical design, this approach tells us little about what a robot can actually accomplish. A surgical robot and a construction robot may share similar kinematic architectures, yet their performance requirements couldn't be more different. Surgery demands high dexterity and gentle touch; construction requires strength and endurance.

The tree of robots addresses this disconnect by starting with the tasks themselves. The researchers analyzed industrial manufacturing processes, drawing from vocational training curricula, trade show demonstrations, factory implementations, and roughly 870 minutes of video footage showing robots in action. They identified 17 core processes grouped into five categories: assembly and disassembly, dispensing, welding and soldering, handling, and processing.

From these processes, the team extracted eight basic operations that robots must perform: go to pose, follow path, establish contact, follow contour, apply material, manipulate workpiece, sense force, and apply force. Each operation comes with quality requirements—how accurately must a position be reached? How precisely must a force be controlled? These requirements translate directly into robot fitness metrics.

Measuring Tactility

The research introduces 25 new metrics focused on what the team calls "tactility fitness"—a robot's ability to sense and control physical interactions with its environment. These metrics sit alongside existing motion performance standards, which measure positioning accuracy and repeatability, to create a comprehensive picture of robot capability.

Force sensing metrics evaluate how accurately a robot can measure external forces. The researchers attached calibrated weights to robot end effectors and recorded the internal force measurements. Accuracy ranged from 0.81 newtons for the Franka Emika robot to 2.73 newtons for the Yuanda Yu+. Force sensing resolution—the smallest change a robot can detect—varied from 0.06 newtons to 4.23 newtons across the tested systems.

Drift over time matters too. A robot might sense forces accurately at first, but if its readings degrade over minutes or hours, it becomes unreliable for sustained work. The team measured drift at intervals of one minute, ten minutes, one hour, and eight hours. Some robots showed drift of just 0.02 newtons over one minute, while others drifted by 0.24 newtons. Over an eight hour workday, drift ranged from 0.05 newtons to 0.43 newtons.

Force control metrics assess how well a robot applies desired forces to surfaces. The researchers used a specialized test rig with a force sensor covered in hard plastic, contacted by a stainless steel sphere attached to the robot. The setup mimics common industrial scenarios like handling printed circuit boards or testing screens. Force control accuracy—how closely the applied force matches the commanded force—ranged from 0.55 newtons to 1.20 newtons among systems with advanced force control.

Overshoot is critical when handling fragile materials. If a robot slams into a surface before settling to the desired force, it risks damaging delicate parts. Measured overshoot forces ranged from 1.75 newtons to 17.55 newtons, and settling times varied from 0.45 seconds to 3.72 seconds.

Safety and Sensitivity

Human safety metrics evaluate whether robots meet current safety standards for collaborative work. The team used a standardized collision testing device to measure forces during impacts with different body parts at various speeds. Conformance to safety thresholds ranged from 22.5 percent to 83.75 percent for transient collisions and 46.25 percent to 100 percent for quasi static contacts.

Contact sensitivity measures how reliably a robot detects and reacts to unexpected collisions. The researchers built a passive pendulum with adjustable mass and contact stiffness, then programmed robots to move toward it at different speeds. Detection success at high speeds (above 250 millimeters per second) varied from 3 percent to 100 percent. At lower speeds, used for careful object search, detection ranged from 0 percent to 100 percent.

Manual maneuverability quantifies how easily a human operator can physically guide a robot by hand—an increasingly important feature for programming and adjustment. The metrics capture the force required to start motion, maintain steady movement, and accelerate the system. Minimum motion force ranged from 0.94 newtons to 4.14 newtons. Guiding force during steady motion varied from 4.38 newtons to 26.28 newtons.

Dramatic Performance Gaps

The measurements revealed striking differences between robots marketed for similar applications. In some cases, the worst performing system required twice the force or showed twice the error of the best. For metrics like force overshoot or sensing drift over one hour, the ratio between best and worst exceeded tenfold. The most extreme gap—a factor of 1,802—appeared in sensing drift over one hour, between the Franka Research 3 and the Universal Robots UR10e.

These differences matter because they directly determine which tasks a robot can reliably perform. A robot with poor force sensing accuracy might struggle with precision assembly. One with high overshoot forces could damage fragile components. A system with low contact sensitivity might fail to detect obstacles, creating safety hazards.

To classify robots based on these performance differences, the team applied statistical analysis. They used the Mann–Whitney–Wilcoxon test to identify significant differences in tactility fitness between each pair of robots. Robots showing no significant difference clustered together. For example, the Franka Emika and Franka Research 3 showed statistically indistinguishable tactility performance. So did the LBR iiwa and Doosan M0609.

Building on these clusters, the researchers used an expectation maximization algorithm to group the 11 tested robots into four distinct categories: industrial manipulators, cobots (collaborative robots), soft robots, and tactile robots. These categories reflect different levels of overall fitness for physical interaction tasks, from basic industrial operations to advanced force sensitive applications.

A Living Database

The tree of robots is designed to grow. The researchers have released an online tool and open source code that allows anyone to view the classification, add new robot measurements, and test where a system fits within the taxonomy. As more robots are tested and more metrics are contributed, the tree gains resolution and utility.

The framework is also extensible to other robot families and application domains. While the current work focuses on industrial manipulators, the same process analysis approach applies elsewhere. The team demonstrates this with a preliminary analysis of care robots, where tasks like shaving or applying medical ultrasound introduce new requirements: perception accuracy, human comfort, communication precision, and fluency of interaction.

Each new application domain will generate its own set of fitness metrics. Energy efficiency, reflex speed, workspace dexterity, and exteroceptive sensing could all become branches of the tree as different robot types are evaluated. The structure accommodates this expansion naturally, with morphology defining the trunk and branches, and fitness defining the leaves where individual robot systems reside.

Why Classification by Fitness Matters

Current robot specifications focus heavily on payload capacity, reach, and speed—attributes that matter for traditional factory automation. But as robots move into more diverse and unstructured environments, these metrics tell an incomplete story. Two robots with identical reach and payload might differ enormously in their ability to handle delicate objects, react to unexpected contacts, or work safely alongside humans.

The tree of robots makes these hidden capabilities explicit and comparable. It provides a common language for researchers, manufacturers, and end users to evaluate robot fitness for specific tasks. A company considering robots for assembly can compare force control accuracy and settling time. A researcher developing new manipulation algorithms can benchmark against standardized metrics. A safety engineer can assess contact detection performance across platforms.

This approach also highlights gaps in current robot design. If most systems in a category score poorly on a particular metric—say, force control bandwidth or impact stability—that signals an opportunity for innovation. Manufacturers can target specific fitness dimensions to differentiate their products. Researchers can focus on capabilities that matter most for emerging applications.

The framework moves robot classification from a primarily descriptive exercise to a predictive and prescriptive one. Rather than simply cataloging what robots are, it assesses what robots can do and guides what they should become. As robotic systems grow more capable and more varied, this shift from form to function becomes essential.

The tree of robots offers a new lens on an increasingly diverse technological landscape. By rooting classification in process based performance rather than morphology alone, it connects robot design to real world outcomes. And by making the framework open and extensible, it invites a global research community to build a shared understanding of robot capability—one measurement at a time.

Credit & Disclaimer: This article is a popular science summary written to make peer-reviewed research accessible to a broad audience. All scientific facts, findings, and conclusions presented here are drawn directly and accurately from the original research paper. Readers are strongly encouraged to consult the full research article for complete data, methodologies, and scientific detail. The article can be accessed through https://doi.org/10.1038/s42256-025-00995-y