Amazon Boosts Self-Driving Machines With New Patent

Amazon wants its AI to understand the real world. 

The company wants to patent “attention-based deep reinforcement learning” for autonomous agents. This system aims to help autonomous agents that were trained using reinforcement learning to perform with more accuracy in unfamiliar environments. 

Though an autonomous system can be trained on a simulation designed to look as realistic as possible, a “model trained using a simulation may perform poorly when there is a mismatch between the simulation and its real-world deployment environment,” Amazon said in its filing. 

To break it down: Amazon’s system uses neural network-based reinforcement learning models that include what it calls “attention layers.” These layers help the model better adapt their knowledge to real environments from the simulated environments in which they were trained. 

These attention mechanisms rank the importance of different features that sensors detect in an environment, and use those ranks to guide decision-making by highlighting the most important factors. For example, if an autonomous vehicle senses a big rock in the middle of the road and a tree by a sidewalk 20 feet to the right, it’ll rank the importance of the rock over that of the tree, and move the vehicle out of harm’s way accordingly. 

It’s not a surprise that Amazon is working on ways to improve autonomous capabilities, especially as it relates to self-driving vehicles. The company owns a subsidiary called Zoox, an autonomous car startup that it purchased in mid-2020 with the goal of helping it “bring their vision of autonomous ride-hailing to reality.” 

Zoox started offering robo-taxi rides in Nevada and California in early 2023, and reportedly grew its headcount 16% between January and July of last year. The company also leased a 219,300-square-foot facility in the San Francisco Bay Area in late September amid plans to boost manufacturing of its vehicles.

But this patent could be applied to more than just autonomous cars: The ecommerce giant could use this technology in robots throughout its dozens of fulfillment centers. Amazon is already working on ways to push automation throughout its facilities, including producing up to 1,000 robots a day at a facility in Massachusetts. Tech like this could be implemented to help those machines be more efficient and accurate. 

Amazon’s been working nonstop on AI initiatives and investments, including introducing two chips, new generative AI services and a new chatbot at its Re:Invent conference in November.

Its patent activity tells a similar story, with this filing adding to an ongoing pile of AI-based tech that the company has sought to make proprietary. Though it’s still fighting for the spotlight in the market among competitors like Microsoft, Google and Nvidia, the company is bolstering its AI prowess in more ways than one.