Search and ranking
Planned (step 4): a search over edits ordered by a learned ranker, measured in calls to the local solver.
Nothing on this page is built yet. It records the plan agreed on Nov 7, 2026; the page will report results as they come.
The search
A node is a formation with a rounded pose. From a node, the candidates are the edits that pass the hold gate; each expansion calls the local solver Λ, which is the expensive part. The search is policy-guided (Levin tree search, or policy-guided heuristic search): a ranker gives a probability π to each candidate edit, nodes are expanded in order of depth divided by the probability of their path, Λ is charged when a node is expanded, and edits the gate rejects get π = 0. For this order, Orseau and Lelis bound the number of expansions by depth / probability of the solution. Duplicate states and a Λ that fails at random break the tree assumption behind that bound; both are open.
The ranker
- Inputs: one token per contact template (type, gap, normal, contact-set length, whether it is on), one per body, and one for the goal; attention across tokens, with learned biases for pairs that are related by containment or share a body or an environment element.
- Output: a score per candidate edit, normalized over the candidates into π.
- Training: supervised, from searches that call Λ without guidance on many scene instances: the policy-guided-search loss on solution paths, plus a cross-entropy on whether each Λ call succeeded.
- Baselines: search without guidance; the gate’s margin as the score; and a learned feasibility classifier with uniform-cost search on −Σ ln P (Hauser, Bretl and Latombe, ICRA 2005).
Scene families
| Family | Scene and goal |
|---|---|
| F1. Step, randomized | step height, box size, friction, starts; goals on the platform; also lowering the box from the platform |
| F2. Staircase | two or three steps of random rise and tread |
| F3. Wall and ledge | topple the box against a wall, or lift it onto a ledge beside it |
Each family with one or two disks, and the motor limit as a parameter (two weak disks still need the environment; two strong ones can lift the box). Validity is computed from the geometry for each family (Issues).
The test: a 2 × 2 table of transfer, holding out an environment family (F3) and holding out the number of disks, with cost measured in Λ calls. A targeted literature search found learned guidance for contact or mode search tested on held-out instances, objects, object counts or robot counts, but not on a held-out family of environments; the closest work counts solver calls with learned feasibility for pick-and-place (Ortiz-Haro et al., ICRA 2023).