74. Robots Working on the Moon and in Warehouses

Engineering & simulation

Researchers used a reinforcement learning model to allow warehouse robots to plan paths in dynamic environments by exploring surroundings and learning from obstacles, enabling decentralized navigation without a central computer.

“this is a reinforcement learning model that allow robots to plan their path in large dynamic environments … They're just exploring their surroundings to be like, I can go here. I shouldn't go there. And they're learning from this process.”

Engineering & simulationrobust cross entropy modeloutperform the current state of the art quite substantially

The team developed a safety-focused reinforcement learning model with specific constraints to ensure robots avoid collisions with people and goods, benchmarking it against state-of-the-art models in a virtual environment.

“this one is called the robust cross entropy model. … to make sure that their model is conscious of safety as like priority number one, they create safety constraints. … they outperform the current state of the art quite substantially.”

Engineering & simulationcontext aware safe reinforcement learning

A reinforcement learning model was created to help robots react to unknown dynamic disturbances like moving people or trash by continuously checking current planning policies against recent sensor data to avoid freezing or crashing.

“the reinforcement learning model they made here, the robot is continuously taking into account its current planning policy … and then incorporating the data for the past few seconds and minutes to keep checking, is this still the right plan?”