CARD links geometric proof certificates to robotic actions

A new research framework separates certified estimator results from the physical transforms robots execute, then tracks how proof support can be acquired and used. Its study reports distinct certificate and navigation results across 404 airborne-survey cases, alongside tests of service selection and graph-computation costs.

CARD, short for “Proof-Carrying Certificate-to-Action Interface for Relative-Measurement Graphs,” is a framework for connecting mathematical certificates in geometric estimation with actions a robot can execute. The work emphasizes that a checker’s result, a physical transform, and the proof support needed to authorize a service are separate objects and should be recorded as such.

The framework represents an action together with its exact proof witness. This lets it distinguish a change in the action from a change or repair in the witness supporting it. The authors derive a condition for when the cost of tracking both is the same as tracking the action alone: a semantically optimal action sequence must have a witness path with zero repair cost.

CARD also models proof acquisition as a time-sensitive decision. Under uniform sampling without replacement, the number of useful services follows a hypergeometric yield. When requests are drawn from the eligible population, each request produces a useful service under the study’s observed live-confirmation condition.

The primary evaluation covers 404 Full16 cases from 20 sorties of an AA15-X500 airborne survey system. All 404 checker-confirmed optimizer points differed from the reference, with median rotation and translation discrepancies of 1.9891 degrees and 21.6322 meters. By contrast, the paired causal-navigation transforms were reference-consistent in all 404 cases; their median discrepancies were 0.0036566 degrees and 0.0104519 meters.

A separate acquisition study used 170 candidates, of which 155 were eligible, and tested five request budgets. The largest gap between the hypergeometric prediction and the observed mean number of services under ordinary random selection was 0.8412. Ordinary matched random selection achieved 0 of 50 target successes, while actionability-matched random selection achieved 50 of 50 and matched the CARD policy.

For downstream evaluation, the authors tested 11 cases from KITTI, EuRoC, and TUM SLAM datasets. Random graph interventions had optimizer wall-time ratios ranging from 1.128 to 67.65 times those of CARD interventions. Solver records attributed the difference to elimination fill, Cholesky factor size, and numeric factorization.

Sources: arXiv cs.AI.

Related articles