Last month, the Massachusetts Institute of Technology and Mecalux SA announced the establishment of the MIT Intelligent Logistics Systems Lab. The laboratory is part of a five-year project to explore the integration of artificial intelligence models in logistics.
The partners said they will focus on increasing the productivity of autonomous mobile robots (AMRs). Using simulation, optimization, and machine learning techniques, the researchers plan to develop a “swarm intelligence” system enabling multiple robots to operate as a single entity, making collective decisions.
“We aim to create a new generation of autonomous robots that learn from human behavior to foster greater collaboration and efficiency in warehouses,” stated Dr. Matthias Winkenbach, director of research at the MIT Center for Transportation & Logistics (CTL) and the new lab.
The researchers will also work on training self-learning AI models. The Intelligent Logistics Systems Lab intends to design systems that can learn from demand patterns and anticipate customer purchasing habits.
Winkenbach replied to the following questions from Automated Warehouse:
Mecalux, MIT algorithms to be hardware-agnostic
Are Mecalux and MIT working on swarm intelligence using Mecalux’s robots or those of other suppliers?
Winkenbach: Our primary use case for now are Mecalux AMRs, but conceptually, our algorithms are hardware-agnostic and could be adapted to other platforms as well.
Will the team’s machine learning systems work across hardware brands? Will they work with different types of robots?
Winkenbach: The algorithms are not hardware-specific per se. Only the parameters that define the physical characteristics of the robot such as speed or size are specific to the hardware. This means that the models we develop could be applied to other robots, but may need to be retrained for robots with very different characteristics.
Intelligent Logistics Systems Lab applies AI to swarms
How will the swarm systems be different from existing fleet management, warehouse orchestration, and management software?
Winkenbach: 1. They will have stronger predictive capabilities, adjusting the task assignment to robots and the robot navigation to the current and expected future traffic/demand in the system.
2. They will optimize globally, meaning across the entire fleet of robots, not on the individual robot level.
3. They will take into account the real-time state of the system, including robot location, remaining battery levels, expected incoming requests, etc., to make comprehensively data-driven decisions rather than myopic rule based ones.
research at MIT CTL. Source: LinkedIn
Are there specific inefficiencies or bottlenecks that the self-learning AI models will address?
Winkenbach: They will capture interdependencies between robots and tasks.
For instance, it may assign a task to a seemingly suboptimal robot — such as one that is farther away — in order to keep another robot available for an incoming high-priority order, or for an order that would allow that robot to connect multiple tasks on an efficient route.
Warehouse systems could learn as they go
Can you describe the approach MIT CTL and Mecalux will be taking to enable warehouse systems to continually improve?
Winkenbach: Learning models are by definition able to adjust to changing environments through continuous retraining. Especially reinforcement learning approaches allow for constant adjustment of model behavior as the environment changes.
How much and what kinds of data are needed to create these systems, and how will it be gathered — by robot, existing systems, or from a higher-level function?
Winkenbach: These model require significant amounts of historical data on tasks/demand to be handled by the robots. Moreover, a solid simulation test bed is required for training and validation.
Will the self-learning system use or support simulations of different robots and warehouse environments?
Winkenbach: A simulation test bed is important to train models on new warehouses layouts, new demand profiles, or new robot variants.
It also allows for safe and cost efficient testing and validation before the algorithms get deployed to the real-world warehouse environment.
Will either of these systems take into account existing or remaining manual processes within the supply chain?
Winkenbach: Existing processes, rules, or algorithms may serve as a starting point for learning based models to start from and improve over.