Robot Learning

Basic concepts

State estimation is the process of determining the state of a system based on observations and a model of the system. In robotics, this often involves using sensors to gather data about the environment and the robot’s own state, and then using algorithms to estimate the robot’s position, velocity, and other relevant variables.

Note

Moravec’s paradox: things that are hard for animals (abstract thinking) are easy for computers, and things that are easy for animals (perception, motor skills) are hard for computers.

Robot learning applies data-driven methods of machine learning to robotics, enabling generalization and adaptability, to solve problems too complex for traditional hand-coded algorithms.

Note

Current north star problem for robot learning: generalist robot policies. Get large robot datasets, train robot foundation models, apply and fine-tune them on any task, any robot, any environment.

Current advances:

• Large datasets: Open-X-Embodiment
• Large models: Octo, Issac GR00T, Google RT-2, pi-0

Imitation learning

Given labeled data , learn . Basically just supervised learning but for actions.

Reinforcement learning

Learn behavior . Assumes data is not i.i.d., actions affect future states.

Policy optimization

Vision-Language-Action (VLA) models

Foundation models for robotics