THE IDEA ON SCREEN
IN THIS FILM 08 CHAPTERS
INTERACTIVE EXPERIMENT
Educational toy modelWHAT CHANGED
TRY THIS
HOW TO READ IT
THE SOURCE PAPER
The animations, toy geometries, and numerical exercises are original explanations. They are not reproductions of the papers’ software or benchmarks. Paper-specific results are explicitly identified.
Download transcript ↓ Download captions ↓THREE QUESTIONS TO CARRY FORWARD
- What is being searched: actions, continuous values, constraints, or programs?
- What evidence makes a proposed behavior feasible?
- Which assumptions connect the model to the physical world?
Read along
THE WHOLE SERIES
Same robot. Different questions.
An interpretive comparison of the five contributions. This is a reading guide, not a performance ranking.
| Paper | Main contribution | What gets reasoned about | Key boundary |
|---|---|---|---|
| Integrated TAMP | A survey of the coupled planning problem | Discrete actions and continuous feasibility | A family of methods, not one solver |
| PDDLStream | A declarative sampler interface and planning algorithms | Plans, stream dependencies, parameter bindings | Sampler contracts and model correctness |
| OWL-TAMP | Language-generated task constraints | Ordering sketches and continuous predicates | A checked constraint may still misstate intent |
| TiPToP | A modular perception-to-action system | Scene grounding, cuTAMP plans, execution | Reported execution has no visual replanning |
| GaP | Agent-generated and rehearsed skill graphs | Program structure and parameters across instances | Known workcell, bounded variation, available skills |
“GaP reinvented TAMP” captures the family resemblance. More precisely, its outer search generates executable programs, while TAMP addresses coupled action and geometric planning. Those programs can themselves call motion planners or learned skills.