AI and the future of port automation
Fred Dorosh Fred Dorosh
May 6 9 min

AI and the future of port automation

Around the world, ports are facing increasing pressure to adapt to rapid TEU growth, both in terms of volume and vessel size. Yet zoning and regulatory constraints often limit the ability of ports to scale berthing and yard space. In the struggle to remain competitive in this fast-paced landscape, terminals are increasingly turning towards automation in order to optimize the use of constrained assets while minimizing vessel, container and truck turnaround times.

In the pursuit of greater terminal efficiency, equipment automation has become a focus area, attracting $10 billion in investments worldwide. Yet there are still drawbacks as these solutions mature: in a recent McKinsey survey, for example, respondents indicated that while operating expenses at automated ports declined by 15 to 35 percent, productivity also fell by 7 to 15 percent.

Meanwhile, the increasing adoption of open port data-sharing platforms in the semi-private and public spheres has opened the door to a whole new field of opportunities: artificial intelligence at the service of back-office automation. By its very nature, AI-augmented decision-making is less costly to implement, much faster to deploy and less disruptive to operations than equipment retooling. At the same time, it has the potential to enhance and surpass efficiency gains offered by the latter.

As a byproduct of automated equipment deployment, we are also witnessing a boom in edge-device networking and telemetrics in the yard. Well leveraged, this groundswell of data could equip ports with a real-time aggregated digital representation of their operational state. Crucially, this information-rich “digital twin” represents an important windfall for notoriously data-hungry AI training processes. The deployment of Internet-of-Things telemetry all the way from intermodal gates to the quayside unlocks new innovation frontiers in responsive AI: reactive anomaly detection, notification and just-in-time planning.

In summary, community information-sharing systems and trends in mechanical automation are favoring data aggregation and availability in ports, paving the way for AI to become the next transformative technology. The question now is how to leverage it.

Intelligent port orchestration

At Element AI, we envision an AI-enabled port ecosystem that efficiently orchestrates the flow of vessels and cargo throughout the port and beyond.

As data-sharing Port Community Systems (PCS) gain adoption, various actors in the ecosystem (port authorities, terminal operators, carriers, beneficial cargo owners, 3PLs and so on) are incentivized to align their digital transformation roadmaps. This allows them to take advantage of new mutually beneficial opportunities to improve efficiency.

While PCS offer an unprecedented ability to share operational data, true end-to-end real-time optimization across logistics partners requires an intelligence sharing platform. Built on top of a PCS, such a platform would provide a standard interface to share insights, predictions, common goals and constraints across collaborative AI actors in an open framework.

Intelligent port orchestration.

Police by Seung Hnam from the Noun Project; Shipping Container by Nikita Kozin from the Noun Project; Cargo Ship by Nikita Kozin from the Noun Project; Storage tank by ProSymbols from the Noun Project; Freight train by Nikita Kozin from the Noun Project; Anchor by bezier master from the Noun Project

Collaborative AI: the network effect

To be achievable, the vision of collaborative AI must be carried out in rational steps backed by compelling business cases at the pace of individual organizations. Each step involves deploying an AI-driven agent which is linked to traditional systems of record and other agents in a network supported by an open intelligent API. In such a network, AI agents:

- individually augment and inform key human decision-makers for a single operational purpose (e.g. Crew Planning), and

- together act in concert to streamline operations under a common goal.

The favoured approach is therefore a progressive deployment of pockets of AI capability that converge towards a collaborative orchestrated system.

By empowering AI agents to negotiate with each other across organizational boundaries, this collaboration could stretch across the port ecosystem. Acting to optimize KPIs dictated by their respective owners’ prerogatives, agents could identify common goals and mutually beneficial opportunities in near real-time:

  • Vessel turnaround time would be improved at the level of the terminal operator by optimizing berth and quay crane allocation using foreknowledge of clustered vessel patterns. These clusters could be proactively broken up through coordination with ocean liners, pilots, tugs and vessel traffic services.
  • Vessel flow would be improved at the level of the port authority by leveraging insights into navigational bottlenecks in estuaries, anchorages and berthing areas. On a longer timescale, this insight may also inform hydraulic infrastructure programs.
  • Container dwell time may be reduced by providing visibility to hinterland transporters, beneficial cargo owners and third-party logistics partners on the state of containers and predicted milestones throughout the logistics chain.

With an AI network-driven approach, KPIs could be combined and rebalanced on a daily basis as business conditions change. In summary, PCS have the potential to evolve from data-sharing platforms into intelligence-sharing platforms.

Deep learning meets operations research

Many aspects of port and terminal processes fall within the well-established field of operations research. For decades, this discipline has provided tools and methods to encode safety, seaworthiness and regulatory compliance that must be adhered to while generating plans that maximize business metrics and KPIs. In theory, this approach can safely optimize the activities of hundreds of vessels, cranes, yard vehicles and staff moving tens of thousands of containers per day.

Today’s automated planning systems, unable to manage the entire scope and complexity of terminal operations, split planning tasks into multiple subproblems to be tackled in isolation. For example, berth allocation may be solved separately from crane allocation, move sequencing, stowage and yard stacking. This leads to lost opportunities: a slightly less optimal usage of berth space may be necessary to maximize stack proximity and achieve an overall more efficient ship turnaround time. Inefficiency emerges at the margins.

Deep learning meets operations research.
Cargo Ship by Nikita Kozin from the Noun Project; Shipping Container by Nikita Kozin from the Noun Project

What if all these problems could be solved concurrently, enabling stacking plans to dovetail with berthing and stowage plans in a globally optimal solution?

Fundamental research initiatives at Element AI are aiming to combine reinforcement learning techniques with traditional optimization methods. By interacting with real-world systems through accelerated simulation, AI agents could rapidly gain years of experience and something approaching intuition about those systems. We believe that context-sensitive heuristics driven by deep learning have the potential to break down the silos that are inherent in traditional planning systems. This in turn will unlock new levels of efficiency in terminal operations without the need for expensive physical automation infrastructure — simply by building a better plan.

Human-friendly AI

As AI systems gain in sophistication and accuracy, a major priority for Element AI and the broader community is to increase the trustworthiness and accessibility of these solutions.

According to a common paradigm for the progression of analytical tools, descriptive systems evolve into predictive systems. Whereas the former may provide insight into operational mechanics, the latter leverages that insight to derive operational foreknowledge. For instance, a model breaking down a terminal’s historical container dwell time by type and origin is descriptive analysis, whereas a predictive analysis might warn of stack congestion in the next 24 hours based on projected inbound volume.

Recently, machine learning AI models have been central in driving this increase of analytical sophistication. But as business tools are transformed, the black box nature of these models may become a stumbling block. A lack of explicit reasoning or justification behind an AI agent’s actions and recommendations can be a barrier to adoption, especially in human-in-the-loop contexts where trust and transparency are crucial.

Increasingly, introspective technologies are being developed to allow various AI models to provide their human collaborators with visibility on their decisions, highlighting the data that influenced them and exposing their level of confidence. Model explainability and other ergonomic innovations will continue to facilitate and promote human-machine collaboration.

Human-friendly AI.
Figure: blue frames give visual feedback on where a vessel-detecting AI agent is focusing its attention. Image generated by Element AI based on image courtesy of Planet Labs, Inc. licensed under CC BY-SA 4.0

Scenario: optimizing asset utilization with collaborative AI

To illustrate how these concepts could work together to enhance decisions in a real-time operational context, let’s use the example of a terminal operations manager interacting with a network of AI agents.

An Operations Manager queries, “How many ground crews will we need to optimize vessel turnaround on berth 8 this afternoon?

  1. A Port Analyst agent employs natural language processing to extract the intent behind human-language queries. Through its network of AI neighbours, it solicits predictions and speculative plans, which are then combined with live and at-rest operational data to produce informed and relevant answers.
  2. A Vessel ETA agent leverages historical service patterns, weather and tidal predictions, financial data and proprietary operational insights to achieve more accurate predictions of vessel arrival.
  3. A Volume Prediction agent uses historical data, economic indicators and Vessel ETAs to provide a unified view of import/export volumes in a three week window, segmented by container type, hinterland transport method and final destination.
  4. A Terminal Orchestration agent leverages Vessel ETAs and Volume Predictions to provide the TOS with a globally optimized berth planning, crane allocation, stowage and yard stacking solution.

The Port Analyst collects results from the above systems and issues an actionable recommendation: “Optimal assignment is three crews, based on vessel arrival, volume predictions, operational costs and historical performance profiles.”

Collaborative AI.
Cargo Ship by Nikita Kozin from the Noun Project; Shipping Container by Nikita Kozin from the Noun Project


The emergence of intelligence-sharing platforms and the use of AI to improve the reach of planning systems offer the potential of major efficiency gains across port ecosystems. Human-in-the-loop innovations such as model explainability and interactive simulation will allow human operators to benefit from AI insight and gain a better understanding of their own organization’s bottlenecks. In port operations as in numerous other business domains, the future will offer AI systems which not only help us understand and predict, but will also actively assist in deciding the best course of action.

As Port Strategy Lead at Element AI, Fred Dorosh’s mandate is to drive and facilitate adoption of the latest advancements in AI research and technology within the port and maritime business communities.