A group of scholars who use AI to reduce traffic congestion applied their principles to warehouse robotics. Many robots move swiftly across a massive robotic warehouse, collecting and transporting things to human workers for packaging and shipping. These warehouses are now commonly integrated into the supply chain across various industries, including e-commerce and car manufacturing.

Preventing collisions

It is challenging to transport 800 robots to their destinations and prevent collisions effectively. The problem is so intricate that even the most advanced path-finding algorithms need help to match the rapid speed of e-commerce or production. These robots can be likened to cars attempting to manoeuvre through a congested urban area. A team of MIT researchers specializing in AI for alleviating traffic congestion utilized concepts from that field to address this issue.

Deep learning model

A deep-learning model was constructed to encode crucial details about the warehouse, such as robots, scheduled paths, duties, and impediments. This model is utilized to forecast the optimal regions within the warehouse to alleviate congestion and enhance efficiency. The warehouse robots are categorised into groups to expedite the decongestion process using conventional methods for robot coordination. Their technology clears robot congestion almost four times faster than a robust random search strategy. This deep learning approach can solve complicated planning problems beyond warehouse operations, such as computer chip design or pipe routing in massive structures. 

Robotic Tetris

The floor of a robotic e-commerce warehouse resembles a fast-paced game of "Tetris" when viewed from above.

A robot retrieves the shelf containing the required item from a specific warehouse region. It transports it to a human operator for picking and packing when a customer order is received. Multiple robots perform this task concurrently, and if two robots' trajectories intersect in the vast warehouse, they could collide. Conventional search-based algorithms prevent collisions by maintaining the path of one robot while recalculating a new trajectory for the other. The issue escalated rapidly due to the increasing number of robots and the possibility of collisions.

Considering relationships

The neural network can efficiently analyze groups of robots by capturing complex interactions among individual robots. For instance, even if one robot starts far from another, their routes may intersect during their journeys.

The technique optimizes computation by encoding restrictions once instead of repeating the process for each subproblem. The remaining 760 robots in the warehouse need to be limited to free up space for a group of 40 robots that are experiencing congestion. One alternate approach is to evaluate all 800 robots in each group at each cycle.

Building interiors

The researchers' method involves analyzing the 800 robots collectively in each iteration rather than separately for each group. They evaluated their method in several simulated scenarios, such as warehouses, areas with random barriers, and maze-like structures resembling building interiors.

Their learning-based technique identifies more effective groups for decongestion, resulting in up to four times faster warehouse decongestion than conventional, non-learning-based alternatives. Despite accounting for the extra processing complexity of running the neural network, their method still resolved the issue 3.5 times more quickly.

Conclusion

In the future, researchers aim to extract straightforward, rule-based insights from their neural model due to the opaque and challenging-to-interpret nature of the neural network's decisions. Less complex, rule-based techniques may be more straightforward to apply and sustain in robotic warehouse environments.

Image source: Unspalsh