Skip to content

Berkeley CS285 Spring 2026 Guide: 25 Lectures, Five Assignments, and the Self-Study Boundary

Aug 22, 2026 1 min
TL;DR Spring 2026 CS185/285 publishes slides for 25 lectures, nine discussion units, five assignments, and starter code; current recordings require bCourses access, while HW4 defaults to an H100, so this is not a zero-cost open course.
Table of Contents
  1. The six-part route
  2. The prerequisite is more than Python
  3. Read the compute labels before committing
  4. What this guide will not reconstruct
  5. References

🌏 中文版

Berkeley CS185/285 Spring 2026 is Sergey Levine's deep reinforcement learning course. Its public page lists slides for 25 lectures, nine discussion units, five assignments, and two default final projects. The starter-code repository is public too. That makes it an A3 material-based course, but not a fully open course.

The missing layer is video. The syllabus places Spring 2026 recordings in the bCourses Media Gallery. The Fall 2023 recordings listed on the official resources page are historical substitutes, not recordings of the 2026 lectures.

This series therefore treats the 2026 slides, sections, and assignments as canonical.

The six-part route

PartOfficial materialGuiding question
1Full mapWhat is public, missing, and prerequisite?
2L1–4, Sections 1–2, HW1Why does imitation learning face distribution shift?
3L5–10, Sections 3–5, HW2–3How do policy gradients, actor-critic, DQN, and SAC connect?
4L11–18, Sections 6–9, HW4–5How do inference, LLM RL, model-based RL, and offline RL connect?
5L19–25How do exploration, theory, multitask learning, and open problems reshape the map?
6All homework and projectsWhat runs on CPU, and what needs a GPU budget?

The prerequisite is more than Python

The official prerequisite is CS189 or equivalent machine-learning preparation, plus familiarity with RL, numerical optimization, and ML. Before starting, write down how states, actions, transitions, rewards, policies, and value functions relate without consulting notes. If that fails, cover the MDP material first.

Read the compute labels before committing

HW1 and HW2 are CPU-first. Most HW3 experiments can run locally, but the handout estimates roughly three GPU hours for each expensive MsPacman or HalfCheetah run. HW4 defaults to a Modal H100 and requires four runs. An HW5 run can take about six hours, before tuning across algorithms and tasks.

The course's compute support applies to enrolled students. A self-learner should run the smallest configuration locally, confirm the pipeline, and only then purchase cloud time.

What this guide will not reconstruct

Without bCourses, an external learner lacks the instructor's spoken explanation, corrections, Ed threads, Gradescope feedback, and office hours. These articles explain the structure connecting public slides and assignments. They do not invent lecture remarks or relabel Fall 2023 video as Spring 2026.

A better completion criterion is an artifact: one derivation note, one implementation that succeeds on a small environment, a result table across seeds, and a failure analysis. That is closer to the course's actual work than merely “watching 25 lectures.”

References