Week navigation
Week 6: Frontier Embodiment
Week 6 -
Frontier Embodiment
V-JEPA, world models, humanoid controllers, dexterous manipulation, and data collection.
V-JEPA 2, V-JEPA 2-AC, LeJEPA
Glossary primer (15 min) JEPA (Joint Embedding Predictive Architecture) — Yann LeCun 2022. An architecture that predicts representations of...
LeWorldModel, ThinkJEPA, world-model–conditioned policies
Glossary primer (12 min) LeWorldModel — Hugging Face's 2025 world model project. JEPA backbone + action conditioning + LeRobot integration....
DreamerV3 — RL in a learned world model
Glossary primer (12 min) DreamerV3 — DeepMind 2023 (Hafner et al.). Universal RL algorithm: same hyperparameters across 150+ tasks. Trains a...
Humanoid whole-body controllers: HumanPlus, OmniH2O, HOVER, ASAP, BeyondMimic
Glossary primer (15 min) WBC (Whole Body Controller) — Policy that controls all joints of a humanoid simultaneously: legs (locomotion) + tor...
Dexterous manipulation: Eureka, DexMimicGen, π0.7 dexterous
Glossary primer (12 min) Dexterous manipulation — Multi finger / multi contact grasping, reorientation, in hand manipulation. The hardest ro...
EgoScale and the data-collection paradigm shift
Glossary primer (10 min) EgoScale — Feb 2026 paper / dataset (Meta + collaborators). Massive egocentric video dataset (10,000+ hours) with s...
Week 6 retro + Track B design + R²-Dreamer notes
Hour 1 — Capstone Track B pre design (40 min) Track B is "World model": train a small JEPA style action conditioned world model on a custom...
What you will know by end of Week 6
- Read the week's source papers without drowning in undefined terms.
- Run the week's core software stack from a fresh clone.
- Explain the week's systems in terms of data, control, and learning loops.