⚡ electrical engineering

Physical Human-Robot Interaction: A Critical Review of Safety Constraints

This paper critically reviews safety constraints in physical human-robot interaction, particularly those defined by ISO/TS 15066, by examining their derivation and assumptions to evaluate their impact on system performance and emphasizing the fundamental role of energy in safety assessment for collaborative industrial robots.

Original authors: Riccardo Zanella, Federico Califano, Stefano Stramigioli

Published 2026-04-03

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Riccardo Zanella, Federico Califano, Stefano Stramigioli

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 human worker and a robot arm are dancing together, passing tools and assembling parts side-by-side. This is the dream of Collaborative Robots (or "cobots"). Unlike old-school robots that live behind cages because they are too heavy and fast to be safe, cobots are designed to work with us.

But here's the catch: How do you make sure the robot doesn't accidentally hurt you if it bumps into you?

This paper is a deep dive into the "rulebook" (specifically a standard called ISO/TS 15066) that tells engineers how to program these robots to be safe. The authors, Riccardo, Federico, and Stefano, act like detectives, breaking down the math to see if the rules are actually working, where they might be too strict (slowing the robot down unnecessarily), and where they might be too loose (risking injury).

Here is the breakdown in simple terms, using some everyday analogies.

1. The Core Problem: The "Bump" Test

Imagine you are walking down the street and someone bumps into you.

If they are a feather, it doesn't hurt.
If they are a heavy backpack, it stings.
If they are a sledgehammer, it breaks bones.

The paper asks: How fast can the robot be moving when it hits you before it becomes a "sledgehammer" instead of a "feather"?

The standard says: "We have a limit on how much force (push) and pressure (squeeze) your skin and bones can take before it hurts." The paper explains how to translate that "pain limit" into a "speed limit" for the robot.

2. The Physics: The "Springy Pillow" Model

To figure out the speed limit, the authors use a simple mental model:

The Robot is a heavy bowling ball.
You (the human) are a springy pillow.
The Collision is the ball hitting the pillow.

When the ball hits the pillow, the pillow squishes (compresses). The paper argues that the danger isn't just how hard the ball hits; it's how much energy gets stored in that squish. If the pillow squishes too hard, it hurts.

The Golden Rule: The robot must carry so little "kinetic energy" (moving energy) that even if it hits you at full speed, the "squish" never exceeds the point where it hurts.

3. The Three Big "Gotchas" (Where the Rules Get Tricky)

The paper points out three major factors that engineers often get wrong or oversimplify:

A. The "Clamped" vs. "Free" Scenario (The Wall vs. The Dance Floor)

Free Scenario: Imagine you are on a dance floor. If the robot bumps your arm, your arm can move backward, absorbing the hit like a shock absorber. This is safer.
Clamped Scenario: Imagine you are pinned against a wall. If the robot bumps your arm, your arm can't move. All the energy goes straight into crushing your arm against the wall. This is much more dangerous.

The Paper's Insight: The current rules sometimes assume you are "free" when you might actually be "clamped." If you are pinned, the robot needs to move much slower. The authors warn that if we don't account for the "pinned" scenario, we might be too optimistic about safety.

B. The "Contact Area" (The Needle vs. The Foam Ball)

Needle: If a robot pushes you with a sharp needle, even a tiny force hurts because the pressure is concentrated on one tiny point.
Foam Ball: If the robot pushes you with a giant foam ball, you can handle a lot more force because it's spread out.

The Paper's Insight: Cobots are designed to be round and smooth (like the foam ball) to spread out the pressure. But if you attach a sharp gripper or tool to the robot, it becomes a needle. The paper argues that safety limits should change based on what the robot is holding, not just the robot itself.

C. The "Effective Mass" (The Heavy Suit vs. The Feather)

The paper explains that a robot doesn't always feel the same weight.

If the robot is reaching out with a long arm, it feels heavy (high inertia).
If it's curled up close to its body, it feels light (low inertia).

The Paper's Insight: Current rules often use a "worst-case" average weight for the robot, which makes the robot move very slowly to be safe. The authors suggest using "smart" math that calculates the robot's weight in real-time based on its pose. This would let the robot move faster when it's safe to do so, without sacrificing safety.

4. The Energy Concept: The "Battery" Analogy

The authors love Energy. They suggest thinking of the robot's safety like a battery.

The human body has a "pain battery." Once the impact drains more energy than the battery can handle, you get hurt.
The robot has a "moving battery."
The Goal: The robot's moving battery must never have enough charge to drain the human's pain battery.

They argue that focusing on Energy is better than just focusing on Speed or Force because it accounts for everything: how heavy the robot is, how fast it's going, and how long the hit lasts.

5. The Big Trade-Off: Safety vs. Speed

The paper concludes with a crucial realization: Safety often kills productivity.

The "Safe" Way: Assume the robot always hits a pinned human with a sharp tool. Result: The robot moves at a snail's pace (0.1 m/s). It's super safe, but the factory is slow.
The "Smart" Way: Use cameras and sensors to see where the human is and what the robot is holding. If the human is far away and the robot is holding a soft sponge, let the robot zoom!

The Verdict: The paper says we need to move away from "dumb" safety (slowing everything down) to "smart" safety (adapting to the situation). If we can build robots that "see" and "think" about the context, we can make them both safe and fast.

Summary in One Sentence

This paper is a guide for engineers to stop guessing how fast a robot can go, and instead use precise physics to calculate exactly how fast it can move based on where it might hit you, how it's hitting you, and whether you are pinned down or free to move—ultimately making our robot coworkers safer and more efficient.

1. Problem Statement

Collaborative robots (cobots) are increasingly deployed in open environments where direct physical contact with humans is possible. While standards like ISO/TS 15066 and ISO 10218-2 provide quantitative safety limits (forces, pressures, velocities) for Power and Force Limiting (PFL) modes, their application often suffers from:

Ambiguity: The standards contain inherent ambiguities regarding the derivation of limits, leading to inconsistent risk assessments.
Over-conservatism: Common interpretations often assume worst-case scenarios (e.g., clamped human body parts, constant robot mass) that unnecessarily degrade robot performance and productivity.
Lack of Context-Awareness: Safety is treated as a static property rather than an emergent characteristic dependent on the specific task, environment, human anatomy, and dynamic interaction conditions.
Misunderstanding of Energy: There is a need to clarify the role of mechanical energy in safety assessment versus traditional force/pressure metrics.

The paper aims to provide a rigorous, self-contained derivation of these safety constraints, critically examine the underlying assumptions, and quantify the trade-offs between safety guarantees and operational efficiency.

2. Methodology

The authors employ a systematic, physics-based approach to reconstruct the safety constraint derivation process found in ISO/TS 15066.

Mechanical Modeling: The human-robot interaction is modeled as a one-dimensional collision between two lumped masses ( $m_R$ for the robot, $m_H$ for the human) connected by a linear elastic spring ( $k$ ) representing tissue stiffness.
Energy Balance Derivation: The paper derives the relationship between the robot's pre-collision kinetic energy and the maximum elastic potential energy the human tissue can store without exceeding pain thresholds ( $F_{max}$ $F_{ma x}$ ).
- It utilizes the conservation of linear momentum and energy to determine the velocity at maximum compression ( $v^*$ ).
- It establishes that the change in kinetic energy ( $\Delta K$ ) must equal the maximum stored elastic energy ( $U_{s,max}$ ).
Critical Analysis of Assumptions: The authors deconstruct the standard's assumptions, specifically:
- Collision Type: Analyzing the difference between fully inelastic (energy dissipated) and general collision models.
- Human Mobility: Distinguishing between unconstrained (human can recoil) and clamped (human is immobilized) scenarios.
- Mass Estimation: Comparing the ISO's constant quasi-static mass approximation ( $m_R^{ISO}$ ) with dynamic, configuration-dependent effective mass ( $m_u(q)$ ).
Numerical Simulation: Using a Franka Emika Panda manipulator, the authors simulate safety limits across 12 ISO-defined body regions under varying conditions (transient vs. quasi-static, clamped vs. unconstrained, constant vs. dynamic mass).

3. Key Contributions

Rigorous Derivation: The paper provides a clear, step-by-step mathematical reconstruction of how pain thresholds are mapped to robot motion constraints (velocity and kinetic energy limits), clarifying the role of the reduced mass and tissue stiffness.
Critical Identification of Assumptions:
- Restitution Independence: The safety condition is determined solely by the state at maximum tissue deformation; the post-impact restitution (elastic vs. inelastic) is irrelevant for the peak force constraint.
- Contact Area Sensitivity: The allowable force scales linearly with contact area. Using nominal $1 cm^2$ areas (as per standard tables) is overly conservative if the actual contact area is larger.
- Clamped vs. Unconstrained: The "clamped" scenario (where the human cannot recoil) imposes significantly stricter limits than the general case. The standard's current treatment of quasi-static limits may not adequately cover this without explicit modeling.
Energy-Centric Perspective: The paper emphasizes mechanical energy as the unifying metric for safety. It argues that while energy-based thresholds are useful for global safety budgets, they must be carefully distinguished from pain thresholds, as pain is non-linear with respect to energy (dependent on strain rate and pressure distribution).
Review of Control Strategies: A concise review of energy-based control methods, including Virtual Energy Tanks, Passivity-based control, and Control Barrier Functions (CBFs), highlighting their ability to enforce safety constraints while maintaining system stability.

4. Key Results

Numerical experiments on the Franka Emika Panda robot revealed significant performance gaps caused by conservative assumptions:

Impact of Human Mobility: Assuming a "clamped" human (immobilized) reduces admissible speeds significantly compared to an unconstrained human. For sensitive regions like the neck and hands, the speed limit drops by up to 82% when assuming clamping vs. free recoil.
Impact of Mass Approximation: Using the ISO's constant mass approximation ( $m_R^{ISO}$ $m_{R}^{I S O}$ ) instead of the configuration-dependent dynamic mass ( $m_u(q)$ $m_{u} (q)$ ) leads to further performance loss.
- For the Face (most sensitive region), using constant mass reduces speed limits to 70% of the dynamic mass baseline.
- For Hands/Fingers, the reduction is even more severe, dropping to 26% of the baseline.
Worst-Case Aggregation: When a system lacks context awareness (cannot identify the body part or if clamping is possible), it must assume the worst case (Face + Clamped + Constant Mass).
- Under these worst-case assumptions, the admissible speed for a collision with the torso or pelvis drops to approximately 6-10% of the optimal speed achievable with context-aware, dynamic mass estimation.
- Conversely, a context-aware system can operate >10x faster than a conservative, fixed-limit system while remaining safe.

5. Significance and Future Directions

Performance vs. Safety Trade-off: The paper demonstrates that "safety" is not a binary state but a spectrum heavily influenced by modeling fidelity. Overly conservative assumptions (static mass, clamped scenarios, worst-case body parts) result in massive inefficiencies in industrial settings.
Need for Context-Aware Systems: Future safety architectures must integrate real-time perception to:
- Identify the specific human body region involved.
- Detect potential clamping scenarios (obstacles).
- Estimate the effective contact area and dynamic robot mass.
Beyond Collision: Current standards focus on transient impacts. The paper highlights the need for new methodologies to address sustained-contact scenarios (e.g., co-manipulation), where repetitive loading or pressure can cause injury without a high-velocity collision.
Integration with AI: There is a call to integrate energy-based safety constraints with AI-driven control (e.g., Reinforcement Learning) to ensure formal safety guarantees while allowing adaptive, high-performance behaviors.

In conclusion, the paper argues that while ISO/TS 15066 provides a necessary baseline, its practical application requires a deeper, physics-based understanding of the interaction dynamics. Moving from static, worst-case assumptions to dynamic, context-aware safety enforcement is essential for unlocking the full potential of collaborative robotics in industry.