{p['tagline']}
{family_eq_block}{p['intuition']}
{p['tagline']}
{p['intuition']}
{p['tagline']}
" f"Pick options above and I'll suggest paradigms.
" return f"The same method can be a policy, an architecture, a world model, a data recipe, and a deployment role. This layered ontology keeps classical control, robot learning, VLAs, sequence models, and world models in one clean frame.
| Method | Control substrate | Objective | Predictive model | Architecture | Data | Role |
|---|
{n_families} projected families and {n_leaves} leaves. These are layout prototypes; the canonical taxonomy remains the layered ontology and ownership audit.
| Family | Owner key | Leaves | Paradigms |
|---|
{tree['subtitle']}
Each tree slices the same methods by a different relationship: training objective, predictive model, architecture, or runtime role. The table underneath marks the important cross-links where one paradigm belongs to more than one branch.
| Paradigm | Primary branch | Connects to | Also connects to | Runtime interpretation |
|---|
The cleanest map is not BC vs RL vs VLA vs world model. It is a stack: controller, objective, predictive model, architecture, data source, and runtime role. Most current papers are new combinations of these layers.
{html_lib.escape(node['plain'])}
The map keeps branches clean by giving every node one primary job. Related ideas appear as links and unlocks, not duplicate children in multiple branches.
{html_lib.escape(stage['subtitle'])}
{html_lib.escape(node['plain'])}
This table is the validation layer behind the map. Ambiguous terms get exactly one primary owner; related ideas are recorded as links rather than duplicated branches.
| Ambiguous item | Primary owner | Ownership test | Allowed links, not duplicate children |
|---|
This tab makes the landscape auditable: every major survey is tied to the branch or validation claim it supports.
| Survey | Source | What it validates | Role in this landscape |
|---|
A clean map does not mean every paper has only one idea. It means each branch owns one kind of claim, while overlaps are represented as typed links.
A branch can be an objective, an architecture, a data regime, a wrapper, or a representation. Some branches are containers for others.
return-to-go, state, previous action, return-to-go, state, previous action.... The model predicts the next action token. If you replace return-to-go with a goal, it becomes goal-conditioned sequence control. If you train only on expert demonstrations without reward tokens, it collapses toward BC. So DT is a receptacle: it contains BC-like supervised learning, offline-RL-style return conditioning, and long-horizon memory in one architecture.
Based on Anirudha Majumdar's March 17, 2026 discussion of whether robotics world models should condition on future actions.
A world model is a learned simulator. Instead of immediately commanding the robot, it predicts what the world would look like next. The robot can then use those predictions to choose an action, train a policy, or evaluate whether a policy is likely to work.
Do actions go into the model as a condition, or come out of the model as part of the generated plan? That single choice changes what data the model can use and what kind of planning it supports.
The model sees the current scene and an instruction such as place the marker in the basket. It imagines a successful video and also decodes the actions needed to produce that video.
Examples: DreamZero, mimic-video, VideoPolicy, Unified Video Action Model, Large Video Planner, Cosmos Policy-style best-of-N planning.
The model receives a candidate action sequence, for example end-effector poses for the next second, and predicts what would happen if the robot executed those actions.
Examples: Dreamer, Veo-Robotics, Ctrl-World, DreamDojo, PlayWorld, World-Gymnast, WorldGym, V-JEPA 2-style latent prediction.
A layered ontology of robot learning — control substrate, objective, world model, architecture, data regime, and deployment role. The tree has {n_leaves} representative leaves across {n_families} projected families, backed by {n_papers}+ papers.