Skip to content

RL Bootcamp — Participant Primer

This is the permanent setup-and-basics companion to the hands-on tutorial track of the Reinforcement Learning Bootcamp. It is deliberately small: install the toolkit, learn the mechanics, arrive ready.

What this site is — and is not

Is: the single installation guide for every edition of the bootcamp, the Conda workflow, a troubleshooting catalogue, the Gymnasium and Stable-Baselines3 basics, and a complete worked example you can run today.

Is not: the tutorial material. The exercises, environments and solutions for a given edition are released to participants at the event. Nothing on this site spoils them.

If you have been handed setup instructions somewhere else — another repository, a PDF, a notebook cell — they are superseded by Installation.


Start here

  • Before you arrive


    The advice that actually matters: when to install, what hardware you need, what to brush up on, and the failure modes that cost people the first hour.

    Before you arrive

  • Install everything


    OS-by-OS instructions and the one-command Conda environment, ending in a smoke test that proves your machine is ready.

    Installation

  • Learn the mechanics


    The environment loop, the training API, and a worked example that trains a real agent end to end — so the live sessions are about ideas, not syntax.

    Basics

  • Something broke?


    The fixes for the problems people actually hit: MuJoCo rendering, ffmpeg, Apple Silicon, Windows long paths, proxies and more.

    Troubleshooting


The 10-minute version

If you do nothing else, do this:

git clone https://github.com/SARL-PLUS/rl-bootcamp-setup.git
cd rl-bootcamp-setup
conda env create -f environment.yml
conda activate rlbootcamp
python scripts/smoke_test.py

If the last command prints Everything works. You are ready for the bootcamp. you are done. If it does not, you have found your problem weeks early instead of during the first session — go to troubleshooting.


What the tutorial track covers

Editions differ, but the skills and the toolkit are stable. Across the sessions you can expect to work with:

Theme What you practise Tools
Tabular RL Value iteration, Q-learning, SARSA on small discrete problems NumPy
Deep RL for control Training continuous-control policies and reading their curves critically Gymnasium, MuJoCo, Stable-Baselines3
Robustness What happens when the deployment world differs from the training world SB3, domain randomisation
Environment design Building an MDP yourself: observations, actions, reward, termination Gymnasium API
Quality assurance Baselines, seeds, ablations — proving an agent has actually learned TensorBoard, evaluation protocol

A theme running through the whole track: reward going up is not evidence of success. You will be asked repeatedly to compare against a baseline and to justify that RL is the right tool at all.


Which parts apply to me?

Edition-specific details

Dates, venue, schedule and registration live on the event website for the edition you are attending, not here. This site is intentionally year-agnostic so it stays correct.


Credits

Written by Leander Grech with Claude Opus 4.8.

Last updated 6 August 2026 · MIT licensed · the version is shown in the header, top right

Every command, figure and number on this site was produced by running it — the training curves and the video come from a real 100,000-step run, not an illustration. Corrections and additions are welcome: open an issue or a pull request on the repository.