Baseball & robotics startups during COVID-19; ROS on Windows 10

Engineering & simulationreinforcement learning, deep neural networks

Pickle Robot Company uses reinforcement learning and deep neural networks to enable their Dill robot to perceive boxes, plan grasps, navigate without collisions, and optimize package stacking (playing Tetris) in loading docks.

“reinforcement learning, deep neural networks, is kind of bread and butter for us to both see the boxes, figure out how to pick them up, figure out how to move them around without hitting stuff, and then most importantly, figure out how to play Tetris with them.”

Engineering & simulationend-to-end trained machine vision system

The team trains a machine vision system end-to-end using the signal of a successful full cycle (pick, scan, sort) to back-propagate errors and improve the robot's ability to pick boxes without obscuring labels or missing sort destinations.

“There's a machine vision system that requires a training data set. Uh, and it's actually trained uh, end to end. So, a successful cycle for us. … So we train on that signal, and that back propagates all the way through to the vision system that helps us figure out how to pick these boxes up.”

Engineering & simulation

A neural network was modified to specifically detect document envelopes versus other parcels, allowing the robot to apply lower vacuum pressure to prevent crinkling, solving a specific edge case identified through manual observation.

“So what do we do? Well, we add to the neural network to detect the difference between a document envelope and other types of parcels. and you see a document envelope, pick a different lower vacuum pressure.”