💻 computer science

Adaptive Collision Sensitivity for Efficient and Safe Human-Robot Collaboration

This paper presents an adaptive collision sensitivity framework that dynamically adjusts motion interruption thresholds based on the effective mass of individual robot links, thereby significantly increasing productivity in human-robot collaboration while maintaining safety standards.

Original authors: Lukas Rustler, Matej Misar, Matej Hoffmann

Published 2026-09-02

📖 8 min read🧠 Deep dive

CC BY 4.0

Original authors: Lukas Rustler, Matej Misar, Matej Hoffmann

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a factory floor where a heavy machine arm and a human worker share the same space, moving around each other to build something together. For decades, safety rules have kept these two worlds apart, usually by building fences or cages around the robot. But as technology improves, we are asking these machines to work side-by-side with people, removing the barriers to create more efficient workflows. The challenge is that when a robot and a human work this close, accidents happen. If the robot bumps into a person, it must stop immediately to prevent injury. However, the current safety rules are so strict that they treat every bump the same way, regardless of how hard the hit actually is. This means a robot might stop for a tiny, harmless tap just as it would for a dangerous collision, causing the work to grind to a halt constantly. The question researchers are now asking is simple: can we make the robot smart enough to know the difference between a harmless nudge and a real threat, so it can keep working without putting anyone in danger?

A team of researchers at the Czech Technical University in Prague has developed a new way to answer this question. They created a system that allows a collaborative robot to decide, in real time, whether a collision is safe enough to ignore or dangerous enough to stop for. Instead of using a single, rigid rule that stops the machine whenever it touches anything, their method calculates the actual force of the impact the moment it happens. They do this by looking at how heavy the part of the robot is that made contact and how fast it was moving. If the math shows the force is low, the robot keeps working. If the force is high, it stops. This approach, tested on real robots and in detailed computer simulations, allows the machines to continue their tasks much more often, boosting productivity by nearly half in some scenarios while still keeping safety within the strict limits set by international standards.

The core of this new method relies on a concept called "effective mass." In simple terms, not every part of a robot arm feels the same when it hits something. A heavy part near the base of the arm carries more weight and momentum than a light part near the tip. Furthermore, the speed at which a specific part is moving changes the force of the impact. The researchers built a system that constantly calculates this specific weight and speed for the exact part of the robot that is touching an obstacle. They call this calculation the "effective mass." By combining this number with the speed of the movement, the robot can estimate the force of the collision before it even finishes the impact. This estimate is then compared against a safety threshold defined by international standards, which specify how much force a human body part, like the back of a hand, can safely absorb.

To test if this idea works in the real world, the team set up a controlled experiment using a six-jointed robot arm known as the UR10e. They covered the robot with a special electronic skin that acts like a sensitive touch sensor, capable of feeling where a collision happens and how hard the pressure is. They also simulated a second type of robot, the KUKA LBR iiwa 7, which uses sensors inside its joints to feel collisions instead of an external skin. The test environment involved a mock task where the robot had to move past several hanging buckets filled with stones. These buckets acted as stand-ins for a human arm, swinging back and forth to create different types of collisions. Some hits were quick and bouncy, like a ball bouncing off a wall, while others were slower and more crushing, like a hand getting caught in a door.

The researchers compared three different ways the robot could react to these hits. The first method was the standard industrial approach: if the robot touched anything, it stopped immediately. This is the safest but least efficient way, as it halts work for even the tiniest contact. The second method used a fixed calculation that assumed a constant weight for the robot, regardless of which part was hitting the object. The third method, which was the team's new invention, used the dynamic "effective mass" calculation to adjust its reaction based on the specific situation. They ran these tests at various speeds, from slow movements to fast ones, and repeated the experiments many times to ensure the results were reliable.

The results showed a clear advantage for the new, adaptive method. When the robot moved at a speed of 600 millimeters per second, the standard approach stopped the machine every single time it touched a bucket, taking an average of 14.5 seconds to complete the task. The new method, however, allowed the robot to continue working through many of the lighter impacts, finishing the same task in just 8 seconds. This represents a productivity increase of over 45 percent. Even at slower speeds, the new method reduced the number of unnecessary stops significantly. In the simulation tests with the second robot, the improvement was similar, with the new method finishing tasks about 30 percent faster than the standard approach. The data showed that the robot was not just guessing; it was correctly identifying that many of the collisions were too weak to cause harm and therefore did not require a stop.

Safety was the primary concern, and the researchers took great care to ensure their system did not compromise it. They verified that their calculations were "conservative," meaning the robot estimated the force to be higher than it actually was. This ensures that if the robot decides to keep moving, the real force is definitely below the dangerous limit. In tests where they measured the actual force of a slow, crushing collision using a certified device, the robot's estimate was significantly higher than the measured reality. For instance, while the device recorded a force of about 98 newtons, the robot's system estimated it to be around 175 newtons. Because the robot's estimate was higher, it acted with a safety margin, stopping when it should have and continuing only when it was truly safe. This conservative approach means the system is unlikely to let a dangerous situation slip through.

The study also highlighted the importance of knowing exactly which part of the robot is involved in a collision. The researchers found that if the system cannot pinpoint the specific link or sensor pad that made contact, it cannot calculate the correct weight and speed, and the safety decision could be wrong. Their method requires the robot to isolate the collision to a specific area, which is why they used the sensitive skin and joint sensors to get precise location data. Without this precision, the system would have to assume the worst-case scenario and stop more often, losing the efficiency gains. The team noted that while their current setup works well for single collisions, handling multiple simultaneous hits would require more complex planning, but for the typical scenarios found in factories, where collisions are rare and usually involve one part of the body, the system is highly effective.

Looking ahead, the researchers acknowledge that their tests used buckets and stones rather than actual human volunteers. While the buckets provided a consistent and safe way to measure forces, a real human arm is more complex, with muscles and joints that can react and pull away from a hit. The researchers believe that because a human arm is softer and more reactive than a bucket of stones, the actual forces experienced by a person would likely be even lower than what the robot measured in the experiment. This suggests that the system is even safer in practice than the tests showed. However, they also point out that safety is not just about physical force; the surprise of a sudden bump can be unsettling for a worker even if it doesn't hurt. Future work will need to address how to make these interactions feel smooth and predictable to the human operator, not just physically safe.

Ultimately, this research offers a practical path forward for the future of human-robot collaboration. By moving away from rigid, one-size-fits-all safety rules and toward a system that understands the physics of each specific interaction, robots can become more capable partners in the workplace. The findings suggest that we do not have to choose between safety and speed. With the right sensors and smart calculations, machines can be both safe enough to work next to people and fast enough to be truly useful, transforming the factory floor from a place of separation into a place of shared work.

Technical Summary: Adaptive Collision Sensitivity for Efficient and Safe Human–Robot Collaboration

Problem Statement
In physical Human–Robot Collaboration (HRC), safety standards such as ISO 10218-2:2025 (formerly ISO/TS 15066) mandate Power and Force Limiting (PFL) regimes. These standards restrict permissible contact forces to ensure human safety. However, current industrial implementations often apply fixed, conservative contact thresholds across the entire robot surface. This "one-size-fits-all" approach leads to unnecessary productivity losses, as robots frequently stop for collisions where the impact force is well within safe limits. Furthermore, standard configurations often fail to distinguish between transient (dynamic) and quasi-static (clamping) collisions, or to account for the varying dynamic properties of different robot links during motion.

Methodology
The authors propose a framework that dynamically adjusts collision sensitivity based on the Effective Mass (EM) of the specific robot link involved in a collision. The core logic involves a five-step collision handling pipeline: detection, isolation (identifying the specific link), identification (estimating force), classification, and reaction.

Force Estimation: Instead of relying solely on raw sensor readings, the system estimates the collision force ( $F_{est}$ ) in real-time using the robot's velocity ( $v$ ) and the Effective Mass ( $m_u$ ) of the colliding link. The estimation follows the equation:
$F_{est} = v \sqrt{k} \sqrt{m_R^{-1} + m_H^{-1}}$
Where $k$ is the stiffness of the human body part, $m_H$ is the human mass, and $m_R$ is the robot's Effective Mass.
Effective Mass Calculation: Unlike static mass approximations, the Effective Mass is computed dynamically based on the robot's kinematic configuration and the direction of impact ( $u$ ) using the Cartesian kinetic energy matrix and the Jacobian.
Adaptive Decision Making: The system compares $F_{est}$ against predefined safety thresholds. If the estimated force is below the threshold, the robot continues its task; if above, it stops.
Strategies Evaluated:
- Factory (FA): Static strategy where any contact triggers an immediate stop (baseline).
- Fixed Mass (FM): Dynamic strategy using a constant mass approximation (Half of Moving Mass) for all links.
- Adaptive Mass (AM): The proposed dynamic strategy using real-time Effective Mass calculations.

Experimental Setup
The framework was validated using two distinct robotic platforms and sensing modalities:

Universal Robots UR10e: Equipped with AIRSKIN, a pressure-sensitive electronic skin covering the robot body. Experiments were conducted in both simulation and the real world.
KUKA LBR iiwa 7: A simulated robot utilizing joint torque sensors for collision detection and isolation, demonstrating the method's applicability beyond skin-based sensing.

The environment simulated a pick-and-place task involving transient collisions with suspended buckets (simulating human limbs) and quasi-static collisions with a certified force-measuring device (simulating clamping).

Key Results

Productivity Gains: The Adaptive Mass (AM) strategy significantly reduced task completion times compared to the standard Factory (FA) approach.
- In real-world UR10e experiments at 600 mm/s, AM was 45% faster (8s vs. 14.5s) than FA.
- At 400 mm/s, the speedup reached over 70%.
- In KUKA simulations, AM achieved a 30% speedup at 600 mm/s.
Collision Handling: The AM strategy resulted in fewer unnecessary stops. For the UR10e at 600 mm/s, the robot stopped only 4 times on average with AM, compared to 8 times with FA.
Safety Validation:
- Quasi-static: Measured forces ( $F_{mes}$ ) during clamping scenarios were significantly lower than the theoretical estimates ( $F_{est}$ ) generated by the AM algorithm (98 N measured vs. 175 N estimated), confirming the method is conservative.
- Transient: Pendulum-based approximations of impact forces also yielded values lower than the AM estimates, further supporting the safety margin.
- Latency: The system latency (contact detection to stop) was approximately 0.22s, well within the 0.5s safety limit for transient contacts.

Significance and Claims
The paper claims to be the first study to adaptively set collision thresholds for collaborative robots, unifying safety compliance with productivity optimization. The authors emphasize that their method allows robots to operate at higher speeds (e.g., 400–600 mm/s) while remaining within ISO force limits, whereas standard calculations would suggest much lower safe velocities (e.g., 130–260 mm/s).

The significance lies in the shift from static, overly conservative safety configurations to a dynamic, physics-based approach that leverages the robot's specific kinematic state. The authors note that while the method requires an accurate dynamic model to compute Effective Mass, it is applicable to any robot capable of collision detection and isolation (via skin, torque sensors, or motor currents).

Limitations and Future Work
The authors acknowledge several constraints:

Model Dependency: The method relies on an accurate dynamic model; unmodeled payloads or discrepancies could compromise safety.
Single Contact Assumption: The current pipeline assumes one contact at a time.
Human Proxy: Experiments used buckets and measuring devices rather than human subjects. While the bucket setup acts as a conservative proxy (lacking human recoil and soft-tissue absorption), the authors state that future work must incorporate human subjects to validate perceived safety and complex biomechanical interactions.
Directional Assumptions: The calculation assumes the impact direction aligns with the link's velocity vector, which may not hold for complex geometries or oblique impacts.

The paper concludes that while the current setup demonstrates feasibility and conservative safety behavior, rigorous safety validation for real-world deployment requires further empirical measurements and high-speed analysis.

More like this