Curriculum guide

Goal

Move from the mechanics of robot learning to the ability to design, implement, and evaluate a small modern system. The sequence is cumulative: later topics reuse the task contract, dataset format, and evaluation habits established in the earliest ones.

What is on each topic page

• An original concept diagram designed for this study group
• A compact method/component comparison
• Key equations with implementation interpretation
• Common failure modes and diagnostics
• A measurable build milestone
• Primary paper and project references where useful

Notes and diagrams are independent study-group material. The source lecture decks are archived under assets/lectures/ and are not published to the site; they retain the copyright and reuse terms of the source material.

Tracks

• Core: read the guide, join the discussion, and complete the minimal build milestone.
• Builder: reproduce a result and submit code plus a short experiment report.
• Presenter: lead one paper discussion using claim → evidence → limitation → next experiment.

Recommended common task

Use one lightweight task across several topics so algorithm differences remain interpretable. Good choices are a simulated reaching task, planar pushing, or a small grid-world for the earliest sessions. Define:

• observation and action spaces;
• success and safety metrics;
• reset and termination rules;
• demonstration and evaluation distributions;
• compute budget and random seeds.

Reading list by topic

Topic	Suggested discussion papers
Introduction to Robot Learning	No required paper; agree on task and evaluation conventions.
Robot Control and MDPs	Simple random search provides a competitive approach to RL; Deep RL Doesn’t Work Yet; Curiosity-driven Exploration
Imitation Learning	Causal Confusion in Imitation Learning; Representation Learning for Visual Imitation; Transporter Networks
Reinforcement Learning I	Evolution Strategies as a Scalable Alternative to RL; Learning Synergies Between Pushing and Grasping; Human-in-the-loop RL for Dexterous Manipulation
Reinforcement Learning II	End-to-End Training of Deep Visuomotor Policies; Eureka; Latent Plans for Task-Agnostic Offline RL
Generative Models for Control	Diffuser; Implicit Behavioral Cloning; Steering Your Diffusion Policy with Latent Space RL
Sequence Models and Transformers	Decision Transformer; ALOHA; Humanoid Locomotion as Next Token Prediction
World Models	Universal Policies via Text-Guided Video Generation; Training Agents Inside Scalable World Models; World Action Models are Zero-shot Policies
Generalist Robot Policies	Language Conditioned Imitation Learning; Gato; π*0.6
Embodied Reasoning and Test-time Scaling	In-Context Imitation Learning; Voyager; Training Strategies for Efficient Embodied Reasoning
Frontiers and Open Problems	A Path Towards Autonomous Machine Intelligence; The Bitter Lesson; Intelligence without Representation

The official course page is the source of truth for its schedule and links.

Capstone checkpoint

By the final session, each team should have a one-page proposal containing a falsifiable question, baseline, intervention, dataset/task, primary metric, compute estimate, ablations, and a stop condition. Prefer a small clean result over an unbounded platform build.