Why Robotics AI is Hard

Why Robot AI Is So Hard: Gradient Descent, Reinforcement Learning, and the Sim-to-Real Loop Lizzy from Wendy Labs introduces Woof the robot dog to show that training robot behaviors is difficult, then explains why AI for robotics is hard by starting with gradient descent and how complex, bumpy optimization landscapes can trap models in local minima. She describes how robotics multiplies this challenge because robots must coordinate many interacting degrees of freedom while dealing with real-world factors like gravity, friction, latency, noise, motor limits, and imperfect sensors, where mistakes have consequences. She introduces reinforcement learning as trial-and-error training driven by rewards, typically performed first in simulation at massive scale. The trained result is packaged as an artifact (policy, weights, configs, rewards, simulation settings, or a containerized Wendy app) and deployed via WendyOS and Wendy Agent, completing an iterative sim-to-real loop of training, deploying, observing, and improving. 00:00 Meet Lizzy and Woof 00:06 Robot Dog Trick Attempts 00:25 Why Robotics AI Is Hard 00:29 Gradient Descent Basics 01:06 Robotics Search Space Explosion 02:00 Reinforcement Learning Intro 02:29 Training in Simulation 02:42 Packaging the Artifact 03:02 Deploying to Real Robots 03:29 The Iterate and Improve Loop 03:39 Wendy’s Mission