Anchored Path Matching weights
Publication status: uploading / incomplete.
This repository stages target/EMA inference weights for Distilling World Action Models for Efficient Bimanual Manipulation via Anchored Path Matching. Formal APM matches the frozen teacher and student instantaneous direction fields at the same detached student-reached state and time. It also uses a restart endpoint anchor and anchor-priority gradient projection.
Audited artifacts
| ID | Storage | Paper status |
|---|---|---|
shared |
bases/stage1-shared/model.safetensors |
exact |
continued |
Stage I base + variants/continued |
exact |
original-no-field |
bases/stage2-original-no-field/model.safetensors |
exact |
original-field-projected |
Stage II base + matching adapter | exact |
restart-no-field |
Stage II base + matching adapter | exact |
apm |
Stage II base + matching adapter | exact |
restart-field-no-projection |
not published | unresolved |
The unresolved row is intentionally absent: no immutable EMA checkpoint in the supplied project tree matched the formal instantaneous-field, Clean-only, routed-topology predicate. A finite-local Clean/Random checkpoint was found but is classified as auxiliary, not relabeled as a paper checkpoint.
Every uploaded variant was materialized with its declared base and matched
against the immutable source EMA using a logical digest over tensor names,
dtypes, shapes, and raw bytes. See manifest.json, audit.json,
checksums.sha256, and verification/.
Guidance contract
- text-condition dropout probability:
0.1 - teacher-target video/action CFG:
5 / 1 - deployment video/action CFG:
5 / 1 - guidance mode: composed at inference
- supported student budgets:
1,2, and4steps
Text-condition dropout is not a guidance scale.
Materialization
The repository is not marked complete yet. Once the matching code revision is public, install it and materialize a variant with:
python tools/materialize.py \
--base bases/stage2-original-no-field/model.safetensors \
--artifact variants/apm \
--output apm-materialized
Only target/EMA inference tensors are included. Online weights, optimizer and scheduler states, trainer recovery state, datasets, simulator assets, and the LingBot-VA teacher are not redistributed.