Purpose of this Paper

Large-scale residential retrofit is central to national decarbonisation strategies, yet deployment consistently underperforms techno-economic projections. Policy frameworks typically prioritise financial incentives, standards, and consumer engagement, assuming that supply chains can absorb accelerated demand. However, empirical evidence from retrofit and electrification markets suggests that logistics coordination, installer sequencing, and governance misalignment frequently constrain uptake despite favourable economics (Rosenow & Lowes, 2020; Rosenow & Eyre, 2016).

This paper argues that supply-chain readiness constitutes a structural determinant of retrofit scalability and that structured supply chain intelligence is required to anticipate and mitigate implementation risk. Rather than treating supply chains as neutral conduits of policy ambition, the study conceptualises them as active mediators of deployment feasibility. Drawing on transition governance scholarship emphasising implementation conditions in shaping socio-technical pathways (Geels et al., 2016; Wilson et al., 2012), the paper examine the operational interdependencies across retrofit supply-chain domains, identify technology-specific delivery vulnerabilities and, highlights governance asymmetries across value-chain positions.

By advancing a diagnostic intelligence framework, the paper aims to inform strategic decision-making in large-scale building transformation programmes.

Design/Methodology/Approach

The study adopts a structured supply-chain intelligence approach grounded in the Supply Chain Operations Reference (SCOR) framework. SCOR’s domain architecture (Plan–Source–Deliver–Return) enables systematic collection and analysis of intelligence across interconnected logistics functions.

Quantitative data were obtained through a practitioner survey covering manufacturers, distributors, retailers, and installers operating within the residential retrofit market. Domain-level performance indices were constructed to measure coordination intensity, perceived fragility, and operational constraints.

To strengthen construct validity and reduce perceptual bias, findings were triangulated with semi-structured industry interviews exploring sequencing complexity, inventory volatility, transport fragility, and warranty-resolution challenges.

The theoretical scope integrates supply-chain governance with energy-transition implementation research. Rather than focusing on optimisation modelling, the approach positions supply-chain intelligence as a diagnostic capability for identifying structural bottlenecks ex ante. This responds to calls for implementation-oriented analysis in energy transitions (IEA, 2023) and extends supply chain intelligence beyond risk monitoring toward policy-feasibility assessment.

The methodological contribution lies in adapting SCOR from firm-level benchmarking to system-level diagnostic intelligence.

Findings

Three core findings emerge. First, retrofit supply chains exhibit tightly coupled operational interdependencies. Warehousing, transport, installer scheduling, and returns handling form a mutually reinforcing system in which disruption in one domain propagates across others. Rapid demand expansion without coordination capacity therefore increases systemic fragility.

Second, scalability varies systematically across technologies. Delivery-sensitive measures - such as heat pumps and high-performance window systems - display greater vulnerability due to transport fragility, sequencing dependency, and higher first-time installation risk. In contrast, warehouse-compatible measures such as insulation align more readily with existing distribution infrastructures. These findings reinforce empirical evidence identifying logistics and coordination barriers in heat pump deployment (Rosenow & Lowes, 2020; Rosenow & Eyre, 2016).

Third, governance asymmetries persist across value-chain positions. Retailers and installers bear customer-facing accountability but lack upstream control over inventory and warranty authorisations. Similar downstream exposure has been documented in retrofit markets (Mahapatra et al., 2013).

Collectively, the findings demonstrate that supply-chain structure actively shapes deployment trajectories. Technology-neutral incentives may unintentionally favour measures with simpler logistics profiles, echoing broader transition research showing that implementation feasibility influences pathway selection (Geels et al., 2017; Wilson et al., 2012).

Value

The paper advances supply chain intelligence by reframing it as a strategic diagnostic instrument for implementation readiness in decarbonisation programmes. It bridges operations management and transition governance scholarship, demonstrating how domain-level intelligence can anticipate bottlenecks before scale-up.

For researchers, the study extends SCOR into policy-level analysis. For policymakers and industry leaders, it provides a structured method for aligning ambition with delivery capacity, thereby reducing systemic risk.

Research Limitations/Implications

The analysis is based on practitioner perceptions within a single national context, which may limit cross-jurisdictional generalisability. Although triangulation mitigates bias, future research could integrate longitudinal logistics data, predictive modelling, and comparative institutional analysis. Further work could also explore digital supply-chain intelligence platforms to enhance real-time coordination.

Practical Implications

Scaling retrofit requires embedding supply-chain intelligence into programme design. Policymakers and industry stakeholders should conduct structured readiness diagnostics prior to demand expansion, identify delivery-sensitive technologies requiring enhanced coordination, align implementation responsibility with operational control and invest in logistics integration, inventory buffering, and sequencing optimisation.

Supply-chain intelligence must therefore move from a reactive monitoring function to a proactive design tool for large-scale transformation programmes.

Acknowledgement:

The authors would like to express their gratitude to RACE for 2030 CRC: Energy Upgrades for Australian Homes (22.H2.S.0365) for funding “WP5 Supply Chain Development” project.

References

Geels, F.W., Sovacool, B.K., Schwanen, T., & Sorrell, S. (2017). The socio-technical dynamics of low-carbon transitions. Joule, 1 (3), 463-479.

IEA (2023). Energy Efficiency 2023. International Energy Agency, Paris.

Mahapatra, K., Gustavsson, L., & Haavik, T., Aabrekk, S., Svendsen, S., Vanhoutteghem, L., Paiho, S., Ala-Juusela, M. (2013). Business models for full service energy renovation of single-family houses in Nordic countries. Applied Energy 112, 1558-1565

Rosenow, J. & Eyre, N. (2016). A post mortem of the Green Deal: Austerity, energy efficiency, and failure in British energy policy. Energy Research & Social Science 21, 141-144.

Rosenow, J., & Lowes, R. (2020). Heating without the hot air: Principles for smart heat electrification. Brussels, Belgium: Regulatory Assistance Project. Available at: https://www.researchgate.net/publication/341654122_Heating_without_the_hot_air_Principles_for_smart_heat_electrification#fullTextFileContent (accessed 19 February 2026).

Wilson, C., Grubler, A., Gallagher, K.S., & Nemet, G.F. (2012). Marginalisation of end-use technologies in energy innovation for climate protection. Nature Clim Change 2, 780–788 (2012). https://doi.org/10.1038/nclimate1576

Purpose of this Paper

The efficiency of modern logistics operations depends on synchronizing routing schedules

with physical cargo loading, a significant computational challenge[1]. Real-world operations

require solving the Integrated 3D Loading Capacitated Vehicle Routing Problem with Time

Windows (3L-CVRPTW; see Figure 1), where route feasibility depends not just on distance

or time but on the geometric constraints of cargo within the vehicle. Sequential heuris-

tics[2] often produce cost-effective yet physically infeasible routes due to packing conflicts,

while exact methods falter under the combinatorial complexity of joint routing and packing

decisions, creating a clear need for learning-based solvers.

We introduce HYDRA-Net, a novel multi-agent framework designed to resolve the com-

plexities of 3L-CVRPTW. The architecture tackles two core challenges: routing under time

windows and 3D packing with vertical stability and containment constraints. Rather than val-

idating packing retrospectively, a practice that triggers costly backtracking, we implement a

synchronized inference loop. In this loop, the routing network (TRANSIT-Net) proposes can-

didate stops that the packing network (STACK-Net) immediately verifies for 3D feasibility.

When constraints are violated, a recovery mechanism masks the invalid move and prompts

the router to select the next-best feasible alternative. This tight coupling ensures that ev-

ery generated plan is both cost-optimized and physically executable, delivering high-quality

logistics solutions in seconds.

Proposed Methodology: The HYDRA-Net Framework

We propose HYDRA-Net, a multi-agent framework for solving 3L-CVRPTW. Recognizing rout-

ing and packing as separate sequence prediction problems, we decompose the 3L-CVRPTW

into two synchronized Transformer[3]-based agents: TRANSIT-Net (routing engine) and

STACK-Net (packing engine). HYDRA-Net orchestrates their interaction to ensure every

decision is both logistically efficient and physically executable (Figure 2).

HYDRA-Net employs two agents, each using Transformer architectures (Figure 3). The first

agent called TRANSIT-Net processes parallel Site, Vehicle, and Time streams augmented

by a topological Graph Attention Network (GAT). Its decoder predicts routing triplets (Next

Node, Vehicle, Arrival Time) via a shared Cross-Attention Bus, using dynamic masking to

enforce capacity and time window constraints. The other agent STACK-Net, manages 3D

loading by processing Box Features against a discretized container grid. A GAT analyzes

the layout to generate precise item-position pairs, while a Physics Engine mask enforces

stability and non-overlap constraints.

To train TRANSIT-Net and STACK-Net, we developed distinct synthetic data generation

frameworks. The serviceable region is modeled as a unit square partitioned into discrete

grids. Customer demand comprises sets of 3 to 5 boxes drawn from three distinct cuboid

types, serviced by a homogeneous fleet with fixed dimensions (ample availability of vehicles

is assumed).

For generating baseline solutions to train TRANSIT-Net and STACK-Net, we use a Hybrid

Metaheuristic (Figure 4) with a Cluster-First, Route-Second, Pack-Verify strategy: cus-

tomers are clustered via K-Means, routed using Time-Window Aware Ant Colony Optimiza-

tion, and each route is validated by a 3D packer enforcing containment and stability con-

straints. Failed packings trigger cluster resizing, forming a feedback loop that guarantees

physically feasible training labels.

Our inference approach uses a Propose-Validate-Commit inference loop (Figure 5) that

tightly couples routing and packing to avoid infeasible solutions. TRANSIT-Net proposes

a candidate triplet (next customer, vehicle, arrival time), which STACK-Net validates for 3D

packing feasibility. Valid moves are committed; invalid ones trigger a recovery mechanism

that masks the failed action and prompts the next-best alternative, ensuring a physically

executable route.

Findings

Our initial experiments utilized 54,000 synthetic instances to train the respective models,

comprising N ∈ {10, 11, 12} delivery locations, each requiring 3 to 5 items. Demand included

cuboids of dimensions 1 × 1 × 1, 2 × 2 × 2, and 3 × 3 × 3, assuming service by a fleet with homo-

geneous vehicles each with container size 10 × 6 × 6. We used 6,000 held out instances for

testing purposes. The principal finding is the successful deployment of a novel multi-agent

decision architecture. By assigning distinct optimization tasks to specialized sub-networks,

TRANSIT-Net for routing and STACK-Net for packing, the system coordinates disparate deci-

sions into a feasible executable solution. Training was supervised by our custom integrated

heuristic, which was allowed 180 seconds to generate high-quality ground truth solutions.

Quantitative analysis confirms that our architecture achieves complete feasibility on un-

seen test cases, strictly adhering to both delivery time windows and complex 3D packing

constraints. Ongoing experiments indicate that HYDRA-Net’s solution quality is progres-

sively approaching the solution quality of the heuristics with every training epoch. This

confirms that the network is effectively learning the complex operational dependencies.

HYDRA-Net generates valid solutions in just 1 or 2 seconds. This dramatic acceleration

demonstrates that the multi-agent coordination mechanism effectively balances physical

executability with high computational speed, validating its potential for real-time deploy-

ment where rapid, constraint-compliant decision-making is critical for logistics efficiency.

Value

Methologically, HYDRA-Net demonstrates that a novel multi-agent framework can effectively

orchestrate collaborative decision-making between specialized networks (routing and pack-

ing agents in this case) to solve high-dimensional, complex problems previously viewed as

intractable for end-to-end learning. By delivering a unified feasible solution (routing and

packing) in almost real time, HYDRA-Net helps enable instant, dynamic fleet management.

The proposed architecture provides a scalable blueprint for training high-performance mod-

els in complex domains where disparate sub-problems must be solved jointly rather than

sequentially, representing a significant leap forward in collaborative artificial intelligence

systems.

Research Limitations/Implications

While HYDRA-Net consistently produces fully feasible solutions for the complex 3L-CVRPTW,

its solution quality is still improving. Future work should investigate self-improvement train-

ing loops to iteratively refine the policy beyond initial supervision, aiming to match or exceed

heuristic performance.

Practical Implications

Logistics operators can use this framework to replace error-prone sequential planning (of

routing followed by packing), with real-time decisions, cutting solution time from minutes

to seconds while ensuring physically feasible routes that prevent loading dock failures and

improve fleet volumetric utilization.

REFERENCES

[1] Rebollo, M., et al. (2024). ”A branch-and-cut algorithm for the 3L-CVRP.” Computers &

Operations Research.

[2] Gendreau, M., et al. (2006). ”Tabu search heuristics for the vehicle routing problem

with two-dimensional loading constraints.” Networks.

[3] Vaswani, A., et al. (2017). Attention is all you need. NeurIPS.