Skip to the content.

Week 01 — Introduction to Robot Learning

← Curriculum · Week 1 of 11 · Next: Control and MDPs →

Complete lecture deck

Complete deck: 44 slides rendered locally. Select any slide to open the full-resolution image.
Lecture 1: Introduction to Robot Learning - slide 1 of 44
Slide 1 of 44
Lecture 1: Introduction to Robot Learning - slide 2 of 44
Slide 2 of 44
Lecture 1: Introduction to Robot Learning - slide 3 of 44
Slide 3 of 44
Lecture 1: Introduction to Robot Learning - slide 4 of 44
Slide 4 of 44
Lecture 1: Introduction to Robot Learning - slide 5 of 44
Slide 5 of 44
Lecture 1: Introduction to Robot Learning - slide 6 of 44
Slide 6 of 44
Lecture 1: Introduction to Robot Learning - slide 7 of 44
Slide 7 of 44
Lecture 1: Introduction to Robot Learning - slide 8 of 44
Slide 8 of 44
Lecture 1: Introduction to Robot Learning - slide 9 of 44
Slide 9 of 44
Lecture 1: Introduction to Robot Learning - slide 10 of 44
Slide 10 of 44
Lecture 1: Introduction to Robot Learning - slide 11 of 44
Slide 11 of 44
Lecture 1: Introduction to Robot Learning - slide 12 of 44
Slide 12 of 44
Lecture 1: Introduction to Robot Learning - slide 13 of 44
Slide 13 of 44
Lecture 1: Introduction to Robot Learning - slide 14 of 44
Slide 14 of 44
Lecture 1: Introduction to Robot Learning - slide 15 of 44
Slide 15 of 44
Lecture 1: Introduction to Robot Learning - slide 16 of 44
Slide 16 of 44
Lecture 1: Introduction to Robot Learning - slide 17 of 44
Slide 17 of 44
Lecture 1: Introduction to Robot Learning - slide 18 of 44
Slide 18 of 44
Lecture 1: Introduction to Robot Learning - slide 19 of 44
Slide 19 of 44
Lecture 1: Introduction to Robot Learning - slide 20 of 44
Slide 20 of 44
Lecture 1: Introduction to Robot Learning - slide 21 of 44
Slide 21 of 44
Lecture 1: Introduction to Robot Learning - slide 22 of 44
Slide 22 of 44
Lecture 1: Introduction to Robot Learning - slide 23 of 44
Slide 23 of 44
Lecture 1: Introduction to Robot Learning - slide 24 of 44
Slide 24 of 44
Lecture 1: Introduction to Robot Learning - slide 25 of 44
Slide 25 of 44
Lecture 1: Introduction to Robot Learning - slide 26 of 44
Slide 26 of 44
Lecture 1: Introduction to Robot Learning - slide 27 of 44
Slide 27 of 44
Lecture 1: Introduction to Robot Learning - slide 28 of 44
Slide 28 of 44
Lecture 1: Introduction to Robot Learning - slide 29 of 44
Slide 29 of 44
Lecture 1: Introduction to Robot Learning - slide 30 of 44
Slide 30 of 44
Lecture 1: Introduction to Robot Learning - slide 31 of 44
Slide 31 of 44
Lecture 1: Introduction to Robot Learning - slide 32 of 44
Slide 32 of 44
Lecture 1: Introduction to Robot Learning - slide 33 of 44
Slide 33 of 44
Lecture 1: Introduction to Robot Learning - slide 34 of 44
Slide 34 of 44
Lecture 1: Introduction to Robot Learning - slide 35 of 44
Slide 35 of 44
Lecture 1: Introduction to Robot Learning - slide 36 of 44
Slide 36 of 44
Lecture 1: Introduction to Robot Learning - slide 37 of 44
Slide 37 of 44
Lecture 1: Introduction to Robot Learning - slide 38 of 44
Slide 38 of 44
Lecture 1: Introduction to Robot Learning - slide 39 of 44
Slide 39 of 44
Lecture 1: Introduction to Robot Learning - slide 40 of 44
Slide 40 of 44
Lecture 1: Introduction to Robot Learning - slide 41 of 44
Slide 41 of 44
Lecture 1: Introduction to Robot Learning - slide 42 of 44
Slide 42 of 44
Lecture 1: Introduction to Robot Learning - slide 43 of 44
Slide 43 of 44
Lecture 1: Introduction to Robot Learning - slide 44 of 44
Slide 44 of 44

Outcomes

Robot learning system loop

Method map

Available signal Natural starting point Central risk
Known dynamics and objective Feedback control or planning Model mismatch
Expert actions Behavior cloning Compounding errors
Online rewards Reinforcement learning Unsafe, expensive exploration
Fixed reward-labeled logs Offline RL Unsupported actions
Mixed vision, language, and action data Generalist/VLA pretraining Dataset and embodiment mismatch

Core notes

Robot learning sits at the intersection of perception, decision-making, and control. A policy maps the information available to the robot—state, observations, language, history, or all four—to an action or action sequence. Learning is useful when writing that mapping by hand is brittle, when demonstrations contain valuable behavior, or when interaction can improve a policy.

The method should follow the supervision available:

Before choosing an algorithm, define the task. Specify observations, actions, control frequency, horizon, reset rules, success criteria, safety limits, and the training/evaluation distributions. “It looks good” is not a metric. Report success rate, time-to-success, violations, and resource cost separately.

The task contract

A reproducible task can be written as

Task = interface + distribution + objective + protocol.

Keep the deployment loop explicit: observe → infer → apply an action → measure → repeat. A model that performs well on stored frames may still fail when its own actions change the next observation.

Failure modes to watch

Concept checks

  1. A robot has thousands of demonstrations but cannot collect new data. Which learning regimes remain feasible, and what distribution-shift risk dominates?
  2. Why can a lower training loss produce a worse closed-loop policy?

Build milestone

Create a one-page task card for a simulated reach, push, or pick task. Include a scripted/random baseline, three seeds, a success metric, a safety metric, and one out-of-distribution split. No learning is required this week.

Evaluation checklist


← Curriculum · Next: Control and MDPs →