Robotics Study Group
A build-first path from robot-learning fundamentals to modern robot foundation models, adapted for the SRA VJTI study group from ETH Zurich’s Robot Learning: From Fundamentals to Foundation Models course.
Each topic gets a written notes page with original study-group diagrams, key equations, implementation notes, and failure modes. Notes are written up as we cover each topic — Imitation Learning is the current worked example of the format.
What we will learn
- Formulate control and decision-making problems as MDPs.
- Build behavior cloning, DAgger, value-based, and policy-gradient agents.
- Understand generative policies, action transformers, and world models.
- Read modern VLA and generalist-policy papers critically.
- Finish with a small, reproducible robot-learning capstone.
Topics
| Topic | Notes | Build milestone |
|---|---|---|
| Introduction to Robot Learning | Open | Define a robot task and evaluation contract |
| Robot Control and MDPs | Open | PID/control baseline plus a tiny MDP solver |
| Imitation Learning | Open | Behavior cloning and DAgger comparison |
| Reinforcement Learning I | Open | Value iteration and DQN baseline |
| Reinforcement Learning II | Open | PPO or SAC with a clean evaluation loop |
| Generative Models for Control | Open | Multimodal action-policy prototype |
| Sequence Models and Transformers | Open | Train an action-sequence transformer |
| World Models | Open | Learn and test a latent dynamics model |
| Generalist Robot Policies | Open | Design a multi-task policy/data interface |
| Embodied Reasoning and Test-time Scaling | Open | Add candidate generation and verification |
| Frontiers and Open Problems | Open | Present a reproducible capstone proposal |
The full reading list, track structure, and recommended session format live in the curriculum guide.
How each session runs
- Before: skim the notes; one member reads each assigned paper.
- Concept check (30 min): explain the topic’s central idea from memory.
- Paper discussion (30 min): identify the claim, evidence, assumptions, and failure modes.
- Build sprint (50 min): implement the topic’s smallest measurable milestone.
- Wrap-up (10 min): record results, blockers, and the owner of the next action.
Definition of done
Every implementation should include a pinned environment, one reproducible command, at least three random seeds where practical, a non-learning baseline, and a small results table. Report failures as carefully as successes.
Existing resources
- Vanilla Policy Gradient tutorial
- Earlier reinforcement-learning notes
- Official ETH course repository for publicly released exercises
Contributing
Use the contribution guide and the session template. Keep notes original, cite sources, and never commit credentials or additional restricted material without documented permission.
Attribution
This is an independent SRA VJTI study-group resource, not the official ETH Zurich course website. See ATTRIBUTION.md for source and reuse details.