Abstract
Dexterous robotic hands are central to the functionality of humanoid robots, enabling adaptive grasping, precise manipulation and classification applications such as egg doneness classification test in service-industry operations. Despite significant advances in sensing technologies and intelligent control systems, many existing dexterous manipulation approaches still struggle to achieve both high precision and adaptability. While some systems show superior performance under similar conditions, applications involving fragile items—such as in egg doneness testing—which demand highly refined and carefully controlled manipulation strategies remain scarce. This paper presents a concise, focused review of humanoid robots, manipulators and the proposed application of dexterous robotic hands and sensors to egg handling tasks, including both grasping and doneness classification. Building on this review, a case study is presented to further illustrate the discussed concepts.
Keywords: AI/ML-driven operations, dexterous hands
1. INTRODUCTION/ PURPOSE/ MOTIVATION
Humanoid Robots and dexterous hands with sophisticated sensor technology and intelligent control systems designed to mimic human form and behaviour have been adopted in high precision manufacturing and classification operations to boost efficiency [1]. This paper investigates its application in fragile object manipulation, using egg doneness classification as a representative case study. The motivation arises from the observation that, despite advances in humanoid robots, accurate and adaptive handling of fragile objects (e.g., eggs) remains challenging, particularly when both precision and perception are required. A key gap addressed is the difficulty in distinguishing raw vs. hard-boiled eggs using robotic systems due to subtle internal dynamics and fragility constraints. This study aims to review extant state-of-the-art manipulators, sensors, and control strategies; identify performance requirements for dexterous manipulation, and to demonstrate the feasibility of a egg classification system through dexterous robotic system.
2. METHODOLOGY
A. Dexterous robotic system setup
The system integrates: a seven-DoF robotic manipulators (e.g., KUKA LBR iiwa, Flexiv Rizon 4S), five-finger dexterous hands, multi-modal sensing with visual-based tactile sensor (thumb), piezoresistive tactile arrays (fingers) and six-axis force/torque sensing measuring:
These interaction forces enable precise real-time full-3D force/ torque feedback [2].
B. Classification strategies
The eggs were subjected to (A) Shake Test: the eggs were shaken at 2–5 Hz, 0.2–0.8 s duration and the signal features from internal motion differences were detected. The physical principle involved in raw eggs was fluid motion causing turbulence and damping, and in hard-boiled eggs, solid interior causing a stable force response. (B) Spin Test where the eggs were rotated (~30 rpm, ~65 mm radius). This method used combined force sensing with computer vision through a convolutional neural network (CNN) employed to recognize its class or type based on camera images dissimilarities. Classification was based on macroscopic rotational stability.
C. Feature extraction
AI models were constructed from two key features derived from the minute residual force signals after shaking before total demise. These were namely Feature A, MeanDecayTau (τ) and Feature B, MeanDamping (δ). In MeanDecayTau, the peak values from each force signal with initial contact force >1N were identified from the vibration envelope curve and extracted using the Hilbert transform. These were fitted using an exponential decay model: 𝐴(𝑡)=𝐴0 𝑒−𝑡/𝜏. Larger τ represents a slower damping, identified as raw eggs. In MeanDamping, from the same set of vibration envelope peaks, a logarithmic decrement between consecutive peaks is computed: 𝛿𝑘=ln[𝐴𝑘/𝐴𝑘+1]. The averaged value is taken as a proxy for damping and used directly as the MeanDamping feature. These features capture post-vibration dynamics for classification.
D. Experimental design
Variables tested include shake frequency, amplitude, and count, Motion directions (vertical, diagonal), Sensor frequency and sensitivity, Algorithm choice and egg type.
3. RESULTS
The difference in rotational behavior between cooked and raw eggs during spinning is enough to be picked up clearly through visual signals.
Sensor performance
In the shake tests it was found that the low-frequency sensors (100 Hz) cannot capture pulse signals reliably and high-frequency sensors (~1000 Hz) are crucial for better accuracy in doneness classification.
Force and sensing requirements
Manipulator performance
KUKA iiwa and Flexiv Rizon were stable for shaking tasks. Realman RM75-6F arm was insufficient for shaking but acceptable for spinning tests. It was found that shake tests require high stiffness and symmetry, while spin tests require rotational control.
Physical behaviour insights
The internal fluid in raw eggs create vortices and damping and energy is absorbed. In cooked eggs, the rigid bodies produce stronger reaction forces.
Spin vs Shake test observations
4. CONCLUSION
The study contributes to the growing field of fragile object manipulation to demonstrate that dexterous robotic systems can effectively classify egg doneness using tactile sensing and dynamic force analysis. The key contributions include: (1) Definition of dynamic features (MeanDecayTau, MeanDamping); (2) Validation of high-frequency sensing (kHz range); (3) Integration of force and vision-based perception and (4) Framework for fragile-object manipulation. Future work will involve a comparison study using multi-modal sensing fusion (vision and tactile), the deep reinforcement learning (DRL), and dynamic tracking in real-time manipulation.