Machine Learning in Tiny Robots
A new research paper by researchers led by Harvard University sheds light into research challenges in developing Machine Learning (ML) solutions for tiny robots and possible future directions. The paper is a very interesting read on how emerging technologies could develop in the future.
As we have seen in previous articles, tiny robots such as the Petoi Bittle have enormous potential in industrial use cases. The Petoi Bittle has major advantages such as a smaller size and much lesser cost (<$300) compared to a similar quadruped robot, Spot from Boston Dynamics (priced starting from $75000). At the same time, making the Bittle ready for industrial use cases means a very different set of research and engineering work due to the availability of low computational power, as well as a limit on the weight of sensors that can be placed on the Bittle. The researchers describe four major challenges in this space, and present possible research directions in each of them. An overview is illustrated in the figure at the top of this article.
Adapting to Size, Weight Area, and Power (SWAP) constraints. One major challenge that tiny robots have when it comes to ML Inference is that they cannot afford to offload data to a cloud service and run algorithms and ML Inference on the cloud. Furthermore, large ML models cannot be deployed on tiny robots due to memory limitations. The authors suggest the use of online learning, where the robot learns about its immediate surroundings directly rather being pre-trained to know some things. Online learning might also be very applicable for tiny robots because these robots do not have to move to much beyond their initial deployed surroundings.