2.1. Requirements for Recyclable Robot Modules

Three properties are required for a robot module to support circular material flows at the EoL:

1. (A)

   Material circularity: Structural components must be manufactured from recyclable or biodegradable materials—to prevent landfilling at the EoL.

2. (B)

   Structural recyclability: The module must be disassembled into separable functional units without tools or destructive operations, thus enabling the non-destructive recovery of high-value components.

3. (C)

   Functional-unit reuse: The recovered units must retain sufficient functional integrity to be redeployed in a replacement module, thereby closing the material loop at the component level.

This study focuses on concepts B and C. Concept A is addressed through material selection (Section 3). Its broader quantification shall be endeavored in future life-cycle assessments. To operationalize Concepts B and C, two interface-level requirements are defined.

• Easy disassembly (R1): Each unit interface must be releasable without tools and without destroying neighboring units, in a well-defined sequence adhering to DfD guidelines ^4,12,3, which include minimal fastener variety, single approach direction, and no adhesives or potting compounds.

• Automatic assembly compatibility (R2): Each interface must be self-aligning (tolerating 1–2 mm placement error from robotic vision) and approachable from a consistent top-down direction, thereby enabling robotic assembly and disassembly—by symmetry. R1 and R2 are evaluated in Sections 3.6 and 5, respectively.

2.2. Design Concept: Dual-Purpose Releasable Interfaces

The unit-assembly architecture decomposes a robot module into functional units (self-contained sub-assemblies, each performing a single function such as actuation, position sensing, thermal management, or electronic control) and connects them exclusively through releasable interfaces that satisfy R1 and R2 simultaneously. A key symmetry is exploited: an interface that self-aligns during insertion guides the pull-off direction during disassembly, and a mechanism that requires no tools to engage requires no release. Therefore, the same three interface types (magnetic docking, magnetic dowels, and mortise-and-tenon joints) are used in both assembly and disassembly. Beyond the interfaces, the architecture adopts a brick-like stacking layout: units are arranged vertically such that assembly proceeds strictly from top to bottom and disassembly proceeds in the reverse order, with no unknown fasteners and no lateral dependencies between units at the same height.

This architecture offers three EoL benefits. First, material stream separation: each unit contains a small, known set of materials that are routable to an appropriate recovery pathway. Second, functional-unit reuse: high-value components (motors, potentiometers, Peltier devices, heat sinks, and PCBs) are individually recoverable without disassembling the units. Third, reduced material diversity: using one thermoplastic for all housing components minimizes contamination during recycling.

The proposed design strategy can be generalized to other robot modules in four steps: functional decomposition, material-stream assignment, single-direction assembly sequencing, and the use of releasable self-aligning interfaces. The specific interface mechanism should be selected based on the target load, required positioning accuracy, and expected reuse cycle number. Therefore, this study contributes beyond the specific prototype geometry; it proposes an intra-module design strategy for reconciling easy disassembly, functional-unit reuse, and robotic-assembly compatibility.

2.3. Related Work and Gap Analysis

Table 1 shows a comparison of representative prior studies against the four design properties targeted in this study: material circularity (M), intra-module disassembly (D), automatic assembly compatibility (AA), and functional-unit reuse (R). Each property is regarded as a binary attribute: it is either explicitly addressed in the system design and validated, or it is not. Modular robot systems (M-TRAN ⁸, SMORES-EP ⁹, EMERGE ¹⁰, and M-Blocks ¹¹) support AA through module-level connectors designed for robotic docking; however, their internal components remain monolithically integrated. Therefore, D and R are not addressed or partially addressed. Biodegradable robot designs ¹⁴ address M through material selection; however, they are intended for single-use disposable applications, and fail to address D, AA, and R. No prior study has explicitly combined these four perspectives for a complete robot actuator module. The present study addresses them with a complete validation of intra-module disassembly and a preliminary validation of automatic assembly compatibility and repeated functional-unit reuse.

3.1. Structural Material

For the structural components of the module (cases and gear housings), the unit-assembly concept requires a material that exhibits the following characteristics: (i) FDM-printable for rapid prototyping of complex geometries; (ii) mechanically adequate at operating stress levels (10–50 MPa); and (iii) recoverable at the EoL, thus satisfying concept A. Conventional FDM materials such as polylactic acid (PLA) and acrylonitrile butadiene styrene (ABS) fulfill (i) and (ii) but offer limited EoL pathways beyond landfill. In this study, cellulose acetate (CA), a biobased thermoplastic (provided by Nature 3D ^a and manufactured using NEQAS OCEAN ^b), was used because it offers reprocessing and biodegradation pathways at the EoL ^15,16,17. The CA product (NEQAS OCEAN ^b) is certified for marine biodegradability. Table 2 summarizes its key properties.

A known challenge of FDM-printed CA is its weak interlayer bonding in the \(Z\)-direction (\(<4\) MPa vs. \({\sim}\)40 MPa in \(XY\)). The mortise-and-tenon interface described in the following subsection mitigates this limitation by routing the primary load path through the \(XY\) cross-section of the tenon. The architecture is material-agnostic: any single-component thermoplastic satisfying (i)–(iii) can substitute for CA without modifying the structural design.

3.2. Unit-Assembly Structure and Releasable Interfaces

Unlike conventional monolithic robot modules optimized for compactness, our module is composed of nine separable functional units categorized into drivetrain (motor, gear train, and position sensor), cooling ¹⁸ (Peltier, heat sink, and fan), and control (PCBs). Each unit can be independently manufactured, replaced, and recovered at the EoL, thus serving all three recyclability concepts. Fig. 2 illustrates the structure and Fig. 3(a) shows photographs of four representative fabricated units.

The units are stacked vertically in a brick-like layout connected through three types of releasable interfaces (Fig. 4): neodymium magnets embedded in CA components for self-aligning, tool-less magnetic connections; iron-rod magnetic dowels inserted into magnetized holes to secure the upper and lower cases; and mortise-and-tenon joints that distribute loads across thicker CA cross-sections, thus mitigating the weak interlayer bonding of FDM-printed CA. The top-down stacking order is specifically designed for compatibility with robotic assembly (Section 5) and tool-free disassembly (Section 3.6).

Although magnetic interfaces are used throughout the module, they are not intended to serve as the primary load-bearing elements. In the assembled state, the drivetrain and motor units are constrained by the nested housing geometry and mechanical limiting features. Consequently, the operational load is primarily transferred through the housing structure instead of directly through the magnetic interfaces. The magnetic interfaces primarily provide positioning, self-alignment, and tool-less releasability, whereas the nested housing constrains relative displacement during operation. This load-path design reduces the dependence on magnetic attraction for structural stability.

However, the prototype is limited to low-load and low-vibration operating conditions, primarily owing to the material and structural implementation of the prototype. The gear train and housing are fabricated from CA, whose strength, fatigue resistance, and vibration stability are lower than those of conventional metal gearboxes and high-performance engineering plastic housings. For high-torque or high-vibration applications, the same unit-assembly concept requires reinforced housing constraints, stronger recyclable structural materials, and additional releasable locking mechanisms.

3.3. Drivetrain Units

The drivetrain comprises a commercial DC motor (FC-280PC), a CA gear train, and a potentiometer (RDC506002) for position feedback. The motor and potentiometer are reusable units that are recoverable through the magnetic interfaces, and the CA gear casing can be recycled as a single-material thermoplastic.

Two four-stage CA gear trains were developed: \({50:1}\) high-torque and \({28:1}\) high-speed variants (Fig. 5). For the high-torque variant, the Lewis bending stress analysis yielded a maximum tooth stress of 14.05 MPa on the critical fourth-stage pinion, with a safety factor \(n_{F} = 2.7\)–\(4.1\) against CA’s allowable stress (38–57 MPa). The Hertz contact pressure (\({\sim}\)7.8 MPa) remains within limits, thus confirming that the CA gears provide adequate strength. Table 3 summarizes the design parameters for both variants.

3.4. Cooling Units

The cooling unit comprises a thermoelectric cooler (Peltier TEC1-04703), an aluminum heat sink, and a fan, all of which are separable from the CA structure via the magnetic interfaces (Figs. 3(b) and (c)). The Peltier unit, heat sink, and fan are individually reusable components, whereas the CA shelf is recyclable. The magnetic interface enables switching between active (with Peltier unit) and passive (without Peltier unit) configurations, thereby extending the service life of the module.

Thermal performance is evaluated in Section 4.

3.5. Control Units

The control unit (Fig. 6) adopts a “double-waffle” stacked PCB architecture in which two boards are coupled exclusively through pin-to-pin connectors—no solder joints exist between boards. The lower custom driver PCB (Fig. 6(b)) integrates a Peltier driver (Monolithic Power MP2329), a dual DC motor driver (Texas Instruments DRV8848), a fan PWM driver, and independent DC-DC voltage regulators providing 12 and 3.3 V rails. The upper board is a commercial Arduino Nano that implements a 50 Hz PID feedback loop using potentiometer position data and issues PWM commands to the DRV8848. A software end-stop limits the shaft rotation to \(\pm270°\) to protect the wiring harness, and the hardware current limit in the DRV8848 prevents stall-induced motor damage, thus directly extending the motor’s reusable service life.

In addition to the intra-module wiring, the custom driver PCB reserves two \(2 \times 3\)-pin connectors—labeled IN and OUT—for intermodule electrical/signal connection. These connectors are intended to transmit power and communication signals between the modules in a daisy-chain configuration. Thus, the control unit supports not only internal functional-unit separation but also external electrical/signal extensibility for multi-module systems. Multi-module communication through these ports is a design provision that must be validated in future multi-module experiments.

For component reuse, each board can be replaced or recovered independently, the pin-to-pin interface is fully releasable, and signal lines (motor PWM, potentiometer feedback, and temperature sensor) are routed through short wire harnesses with polarized micro-connectors for error-free reconnection after disassembly. The PID gains (\(K_p\), \(K_i\), and \(K_d\)) are stored in the firmware and can be re-uploaded to a replacement Arduino Nano without hardware modifications.

3.6. Easy Disassembly and Exchangeability of Units

This subsection evaluates requirement R1 based on the tool-free disassembly of the module. The functional performance after reassembly is validated in Section 4, thereby confirming unit exchangeability.

Table 4 details the disassembly sequence. Beginning from the topmost unit, eight pull or slide operations—all tool-free—are sufficient to separate the module into clean material streams. Full disassembly was completed within 90 s by a practiced operator; the high-priority motor unit (primary reuse target) is accessible after Step 6.

Figure 7 shows the photographs corresponding to the eight sequential steps. No unit was damaged during disassembly, and repeated-use functional performance is evaluated in Section 4.2.

4.1. Drivetrain-Control Performance

Experiments were conducted to evaluate the drivetrain under step inputs at two load levels (0.05 and 0.10 Nm). Fig. 8 shows the representative response at 0.05 Nm. Across the tested commands and loads, the drivetrain achieved settling times of 0.61–3.24 s, steady-state errors between \(-4.17°\) and \(4.00°\), and overshoots of less than 13.5%. Although the settling time increased slightly at higher loads owing to the gear-tooth compliance and static friction, the system maintains stable closed-loop positioning across the full range. These results confirm that the recyclable CA-based drivetrain does not compromise the control performance in the intended operating regime.

4.2. Multi-Cycle Reuse Evaluation

To provide preliminary evidence for the repeated reuse of the functional unit, the same module was repeatedly assembled and disassembled for 120 cycles. Control-performance tests (Fig. 9) were conducted after every 20 cycles under a 0.05 Nm load. At each cycle, nine step responses were recorded using command amplitudes from 50° to 250°: two 100° steps, one 150° step, four 50° steps, one 250° step, and one 200° step. The rise time, settling time, overshoot, and steady-state error were calculated for each step response and then averaged over the tested commands. Fig. 9 shows the evolution of these metrics as a function of the number of assembly-disassembly cycles. A slight degradation was observed after 100–120 cycles; however, the resulting positioning error remained within the control tolerance required for the intended low-load operating regime. The slight performance decline is likely due to minor gear wear in the gear train unit, which is the most mechanically stressed component; however, it did not significantly affect the overall functionality.

4.3. Cooling Performance

COMSOL transient thermal simulations (180 s, motor heat generation 2.5 W, ambient 20 °C) confirmed that active cooling reduced the motor temperature to \({\sim}\)20 °C, whereas passive cooling only reached \({\sim}\)28.5 °C (Fig. 10). Actual experiments (12 V, 0.1 Nm load, 24 °C ambient) validated these results: the active cooling system cooled the motor to room temperature within 200 s, as compared with \(>3000\) s for passive cooling (Fig. 11). Taken together, the drivetrain, multi-cycle reuse, and cooling evaluations confirm that unit exchangeability does not degrade functional performance.

5.1. System Overview

The system (Fig. 12) comprises a Dobot Magician robot arm (4 DOF), an Intel RealSense D435i RGB-D camera, and a PC operating ROS Noetic. ArUco markers provide camera–robot calibration. The workflow is as follows: (i) the camera captures component images; (ii) a YOLOv8-based [19, 20, c] network detects components and estimates two-dimensional (2D) keypoints; (iii) keypoints are aligned to class-specific templates via rigid-body fitting to obtain planar poses \((u,v,\theta)\); (iv) assembly trajectories are planned and executed.

5.2. Component Recognition and Pose Estimation

We adopted a YOLOv8-pose architecture [19, 20, c] that simultaneously predicts the bounding boxes, class labels, and 2D keypoints for each detected instance. The planar pose \((u,v,\theta)\) was recovered by fitting predicted keypoints to a class-specific geometric template via least-squares rigid-body alignment; for near-symmetric components, a symmetry-breaking keypoint (e.g., a unique notch pattern) was included in the template to resolve orientation ambiguity.

The model (YOLOv8s-pose, fine-tuned from ImageNet pretrained weights) was trained for 50 epochs (batch 32, image size \(640 \times 640\)) on a custom dataset captured five component classes (Peltier unit, motor unit, heat sink, gear train unit, and temperature sensor) under representative illumination, partial occlusions, and surface reflections. The dataset included the gear-train unit as it is one of the nine functional units in the complete module and is required for future extension to full nine-unit assembly, although it was not manipulated in the present four-unit experiment. Data augmentation included random flips, hue-saturation-value jitter, and mosaic compositing (disabled for the final 10 epochs) to improve robustness. Table 5 lists the overall detection and keypoint metrics for the held-out validation set at the final epoch. The mean keypoint localization error in the pixel space translates to 1–2 mm positional and 2°–5° orientation errors in the robot base-frame coordinates (Fig. 13).

5.3. Assembly Experiments and Results

Components were randomly placed on a table, and the robot arm assembled them in a predefined sequence. The system successfully assembled the temperature sensor, motor, Peltier unit, and heat-sink unit (Fig. 14), thus achieving an average cycle time of 40 s per module. Although the perception dataset recognizes five unit classes, the current experiment manipulates only the four representative units above; the gear-train unit is included in the dataset to support full nine-unit assembly in the future, but does not constitute the demonstrated four-unit sequence. The magnetic self-aligning interfaces proved to be critical; in cases where vision-based placement was slightly off-target (1–2 mm), the magnetic attraction corrected the final alignment. This indicates that the proposed releasable interfaces can tolerate the positioning uncertainty of a simple fixed-camera with a 4-DOF assembly setup, which approximates the hardware constraints of common factory SCARA-style automation. For the mortise-and-tenon joints (\(<1\) mm tolerance), vision accuracy alone can be insufficient. The 4-DOF constraint limited assembly to top-down operations.

Therefore, the current experiment should be interpreted as a proof-of-concept validation of interface-level assembly compatibility instead of as a production-yield benchmark. Unsuccessful trials were primarily caused by residual orientation error (\(> 5 °\)) after pose estimation; the present magnetic self-alignment compensates for translational errors (\(<2\) mm) but provides limited passive correction for rotational misalignment. This result suggests two system-level improvement directions: redesigning the magnetic interfaces with rotational self-indexing features and replacing/supplementing the eye-to-hand camera with an eye-in-hand configuration to provide localized visual feedback during the final insertion.

These results support the compatibility of the proposed releasable interfaces with robotic manipulation at the interface level. Because the same interfaces are designed to be released in the reverse direction, the results suggest a feasible pathway for future automated disassembly. However, direct validation of robotic disassembly remains a topic for future investigation.