ActionABI: Turning Undocumented Action Tensors into Verified Converters or Honest Refusals

Abstract

Robot learning datasets and deployments routinely carry action tensors whose meaning—which channel drives which degree of freedom, its sign, scale, whether targets are absolute or delta, in which frame, with what pipeline lag, and whether the gripper is inverted—is undocumented, untrusted, or missing. Existing tools either check declared metadata against a specification or define new schemas going forward; neither retrofits legacy or adversarial data carrying no metadata at all. We present ActionABI, a C++20 evidence-first CLI that scores a finite, declared action-interface grammar against timestamped trajectories, retains a calibrated observational-equivalence set instead of breaking ties, and emits a converter only when every required field is supported by evidence—abstaining rather than forcing a unique answer. On a 100-case synthetic ambiguity gate, ActionABI eliminates all 25 false unique certifications a forced-argmin baseline makes, at full coverage. Against ground-truth latent contracts from real ManiSkill GPU simulation (90 traces, 45 contracts), structural fields (permutation, sign) recover to 0.80/0.84 overall after a lag-observable fix, while scale is confounded by the physical controller, not the grammar. A 35-dataset Hugging Face Hub audit scores recovered conventions against documented ones over 139 field-labels with 0 contradictions and 0 false uniques, finding that “OXE” is not one action convention and that ≥ 5 gripper-polarity conventions coexist. Three measurement-driven defects were found, root-caused, fixed, and regression-tested; ActionABI’s C++ core runs load-bearing in a sibling adaptation benchmark at 5.7 × 10⁻¹⁴ parity and a ~13× CPU-regime speedup, while its CUDA path honestly fails a preregistered 5× gate. We make no “first” claims and no hardware claims; every number traces to a committed report.