💻 computer science

Real-Time Safety Evaluation of Human Arm Operations Using a Wrist-Mounted IMU with PSM System

This paper presents a novel real-time safety monitoring system for human-robot collaborative manufacturing that utilizes a wrist-mounted IMU integrated with a predictive safety model based on a wrist-optimized spring-damper-mass and impedance-based probabilistic assessment, demonstrating robust performance across diverse manufacturing tasks.

Original authors: Musab Zubair Inamdar, Alhusain Al Hadrami, Seyed Amir Tafrishi

Published 2026-06-26

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Musab Zubair Inamdar, Alhusain Al Hadrami, Seyed Amir Tafrishi

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 are working side-by-side, passing tools and parts back and forth. Usually, safety systems focus entirely on the robot: "Is the robot moving too fast? Is it too close to the human?" But this paper asks a different question: "Is the human's arm moving in a controlled, predictable way, or are they flailing, jerking, or making sudden, erratic moves?"

To answer this, the researchers created a "smart watch" system (though it's actually a sensor strapped to the wrist) that acts like a predictive coach.

Here is how the system works, broken down into simple concepts:

1. The "Spring-Loaded" Coach

The core idea is based on a Predictive Safety Model (PSM). Think of the human arm not as a complex biological machine, but as a simple spring-and-damper system (like a door closer or a shock absorber on a car).

The Theory: When a human does a repetitive task (like screwing in a bolt), their arm naturally follows a smooth, rhythmic path, much like a pendulum swinging back and forth.
The Prediction: The computer uses this "spring" model to guess where the wrist should be in the next split second if the movement is smooth and controlled.

2. The "Reality Check" Sensor

The worker wears a small sensor (an IMU) on their wrist. This sensor is like a high-speed referee that constantly shouts out the wrist's actual position and speed.

The Comparison: The system constantly compares the Coach's Prediction (where the arm should be) with the Sensor's Reality (where the arm actually is).
The Result: If the human is moving smoothly, the prediction and reality match up perfectly. If the human jerks their hand, overshoots a target, or makes a sudden, panicked correction, the two diverge.

3. The "Traffic Light" System

The paper introduces a safety score that acts like a traffic light:

Green (Safe): The arm is moving smoothly, matching the "spring" model.
Yellow (Caution): The arm is starting to drift or move a bit erratically.
Red (Unsafe-like): The arm is moving in a chaotic, irregular way (like a sudden reversal or a wild swing).

4. What They Tested

The researchers tested this on three common factory tasks:

Fastening: Using a hand tool to screw things in.
Visual Inspection: Looking closely at an object while moving the wrist.
Pick-and-Place: Picking up an item and putting it somewhere else.

They asked volunteers to do these tasks in two ways:

Controlled: Moving carefully and smoothly.
Irregular: Deliberately moving poorly (jerking, overshooting, stopping and starting suddenly) to simulate what might happen if someone is tired, distracted, or rushing.

5. The Key Discovery: Speed vs. Position

One of the most interesting findings is about what the sensor should pay attention to.

The Problem: Wrist sensors can sometimes get a little "drifty" regarding the exact angle (position) of the wrist. It's like a compass that slowly spins off course.
The Solution: The researchers found that looking at how fast the wrist is changing speed (velocity) is a much more reliable way to spot danger than looking at the exact angle.
The Analogy: Imagine driving a car. If you are slightly off-center in your lane (position error), you might be fine. But if you are suddenly slamming the brakes and swerving (velocity error), that's an immediate danger. The system works best when it focuses on the "swerving" rather than the "lane position."

Summary

This paper is a "proof of concept." It doesn't claim to have solved all factory safety issues yet. Instead, it demonstrates that a simple, lightweight sensor on the wrist, combined with a smart "spring" model, can successfully tell the difference between a human moving smoothly and a human moving in a chaotic, potentially dangerous way.

It proves that we can monitor human safety in real-time by watching for irregular jerks and sudden changes in speed, rather than just waiting for the robot to get too close.

Technical Summary: Real-Time Safety Evaluation of Human Arm Operations Using a Wrist-Mounted IMU with PSM System

Problem Statement
While collaborative robotics has enhanced manufacturing flexibility, safety remains a critical challenge as human-robot proximity increases. Current industrial safety standards primarily rely on conservative workspace safeguards and robot-side constraints (e.g., speed limits, separation distances). However, these strategies do not directly quantify the quality, regularity, or control of the human operator's own motion. In collaborative cells, human behavior is highly task-dependent and can become irregular or hazardous due to fatigue, distraction, or rushed operations. Existing literature focuses heavily on robot motion prediction and collision avoidance, while the use of wearable inertial sensing as a continuous, wrist-level safety evaluator for collaborative manufacturing remains underexplored. There is a specific need for a compact, interpretable safety metric derived directly from wearable motion data without requiring full workspace instrumentation.

Methodology
The paper proposes a wrist-specific adaptation of the Predictive Safety Model (PSM), originally formulated for upper-body motion, tailored for collaborative manufacturing tasks. The system architecture consists of three main components:

Sensing and Data Acquisition: A BNO055 Inertial Measurement Unit (IMU) is mounted on the operator's wrist. It provides fused orientation estimates and angular velocity measurements. The system operates on a filtered discrete-time data stream (20 Hz) to suppress noise while preserving task-relevant dynamics.
Predictive Safety Model (PSM): The core of the method is a reduced spring-damper-mass pendulum model. Unlike the full upper-body formulation, this version is tuned for wrist and forearm kinematics with specific effective parameters: an arm mass of 2 kg, a pendulum length of 0.2 m, and a base radius of 0.25 m. The model generates a "task-consistent safe reference motion" based on a reduced state space defined by a gravity-referenced angular coordinate ( $\theta_g$ ) and angular speed magnitude ( $\omega$ ).
Safety Evaluation Pipeline:
- Probability Map: A task-specific probability map is constructed from recorded trials, mapping the reduced state $(\theta_g, \omega)$ to empirical likelihoods. High-probability regions represent regular, repeatable motion, while low-probability regions indicate irregularity.
- Online Prediction: In real-time, the system queries the local probability neighborhood to determine the most likely safe continuation of the motion and propagates the PSM to obtain a predicted safe state.
- Impedance-Inspired Safety Index: The system computes the discrepancy between the measured IMU motion and the PSM prediction in the time domain. An impedance-inspired normalized error metric ( $e_{imp}$ ) is calculated by fusing orientation error ( $e_\theta$ ) and angular-velocity error ( $e_\omega$ ).
- Weighting Strategies: Two weighting cases were tested: "position-priority" (0.6 orientation, 0.4 velocity) and "velocity-priority" (0.4 orientation, 0.6 velocity). The authors hypothesize that velocity-priority is more stable for wrist-mounted IMUs, which are often more robust to short-term velocity changes than to drift in absolute orientation.
- Classification: The safety level is classified as safe, caution, or unsafe-like based on empirical thresholds ( $\epsilon_s = 0.15$ , $\epsilon_c = 0.25$ ) derived from controlled vs. irregular executions.
- Offline Baseline: A spectral error indicator (frequency-domain analysis) is retained as an offline comparative baseline to validate the time-domain results.

Key Contributions

Wrist-Specific PSM Adaptation: The paper adapts the spring-damper-mass Predictive Safety Model specifically for wrist and forearm motions relevant to manufacturing, distinct from previous upper-body formulations.
Real-Time Evaluation Pipeline: A self-contained pipeline that maps wrist motion to a probabilistic safe-reference model and computes safety from orientation and angular-velocity deviations without requiring full-body estimation.
Impedance-Inspired Safety Index: The introduction of a time-domain safety metric that addresses the limitations of relying solely on frequency-domain indicators for online assessment.
Experimental Validation: Validation across three representative manual tasks: fastening with a hand tool, visual inspection, and pick-and-place.

Results
The study was validated through 18 task trials involving five participants. Each task included controlled, nominal, and deliberately irregular (hazardous-like) executions.

Separation of Motion Types: The results demonstrated a clear separation between controlled motions and deliberately irregular motions across all three tasks. Irregular motions (characterized by abrupt reversals, overshoots, and inconsistent timing) consistently produced higher error values.
Velocity vs. Position Weighting: The velocity-priority weighting ( $P^*=0.4, D^*=0.6$ ) proved more stable and effective for online assessment than position-priority weighting. The data indicated that hazardous-like behavior is manifested more consistently through temporal irregularity and abrupt corrective actions (angular velocity errors) than through large absolute orientation offsets.
Metric Performance: The impedance-inspired score showed a clearer separation in central tendency between safe and unsafe classes compared to the frequency-based score. While pick-and-place tasks generated higher absolute safety scores due to their broader motion envelope, all tasks maintained distinct safe–unsafe separations.
Thresholds: In the tested dataset, controlled operations generally remained below the safe threshold ( $\epsilon_s = 0.15$ ), while irregular events consistently crossed the caution threshold ( $\epsilon_c = 0.25$ ).

Significance and Claims
The authors position this work as a proof-of-concept validation rather than a population-scale study. The primary significance lies in demonstrating that a lightweight wrist sensor combined with an interpretable predictive model can provide useful real-time safety information for human-robot collaborative workcells.

The paper claims that:

A wrist-mounted IMU can effectively serve as a continuous safety evaluator without instrumenting the full workspace.
Temporal irregularity in wrist motion is a more reliable indicator of potential safety risks than absolute orientation deviation alone.
The proposed framework offers a practical extension of the PSM approach, making it directly applicable to human-robot collaborative tasks through a reduced, wrist-specific formulation.

The authors explicitly state that the reported thresholds are empirical operating points specific to the tested tasks and dataset, not universal safety limits. Future work is identified as necessary for larger multi-participant validation, task-adaptive threshold selection, and comparison with other real-time monitoring approaches.