Latent Dynamics · 2026-11-05
Detecting Silent Action-Interface Mismatch with Calibrated Relational World Models
Rui Zhu, Xurui Li, Haixu Tang
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
The authors study offline detection and closed-set localization of silent action-interface mismatches (e.g., axis permutations, sign inversions, rotation unit scaling, gripper polarity). Clean held-out transitions calibrate alarms and a fixed inverse vocabulary enables localization. On ten LIBERO Spatial tasks (432 source episodes), at 2.5% matched clean FPR, recall is 46.1% (ridge), 61.0% (four-component joint GMM), and 65.7% (exact conditional). Joint scoring localizes higher (95.8% vs 94.6%).
Why it matters
They treat interface semantics mismatch as a contract-check problem using learned state–action–effect relations, with calibration done on clean transitions. The results separate detection recall from localization accuracy and show conditioning of the same fitted mixture improves recall, while joint scoring improves localization at the matched empirical false-positive rate.
Method
- Learn state–action–effect relations and use them as offline “contracts” for declared inverse transformations after calibrating alarm thresholds on clean held-out transitions.
- Evaluate three relational scores: ridge forward residual, a four-component full-covariance joint Gaussian mixture score, and the exact conditional of the same fitted mixture to isolate input-marginal contributions.
- For an alarm, rank inverse candidates from a fixed vocabulary using the lowest relational score on (s_t, U_j(a), s_{t+1}) and report closed-set Top-1 localization across valid fault rows.
Limitation
They emphasize these are retrospective diagnostics on recorded transitions: replay uses the recorded next state produced by the clean action, and they do not measure task success or safe correction; they also note no alarm cannot certify a correct interface and they provide no cross-robot or physical closed-loop safety/repair guarantee.
Abstract (from arXiv)
A robot action can satisfy shape and range constraints yet carry the wrong semantics: axes may be permuted, signs inverted, rotations rescaled, or gripper polarity reversed. We study learned state-action-effect relations as offline contracts for detecting these mismatches and ranking their likely fields. Clean held-out transitions calibrate alarms; a fixed inverse vocabulary supports closed-set localization. On ten LIBERO Spatial tasks comprising 432 source episodes, a capacity-controlled comparison at 2.5% empirical clean false-positive rate gives 46.1% recall for a ridge forward residual, 61.0% for a four-component joint Gaussian mixture, and 65.7% for the exact conditional of that same mixture. The joint score retains a 14.9 percentage-point advantage over ridge, while conditioning improves recall by a further 4.7 points. Joint scoring instead has higher localization: 95.8% versus 94.6%. These matched-rate comparisons are retrospective diagnostics, not deployment calibration guarantees. Earlier Object/Goal results broaden tasks within the LIBERO convention; DROID, all three HYDRA tasks, short temporal faults, small offsets, exact repair, and simulated mitigation fail their respective gates, while xArm contributes one task only. The evidence supports a bounded offline diagnostic and a detection-localization tradeoff, rather than joint-density superiority or cross-robot safety.
Related papers
- Acting Through Registers
- Diagnosing and Recovering from Observation-Space Shift at Long-Horizon Skill Seams
- Sensorimotor Alignment: A Platonic Proxy for Embodied Policy Success
- UniWALM: Unified World Action Language Model for Embodied Robot Control
- Spa-Bench: A Real-Robot Benchmark for Spatial Grounding in Vision-Language-Action Models
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First public release: 2026-11-04 · Indexed by Pub-AI: 2026-11-05 UTC
Detecting Silent Action-Interface Mismatch with Calibrated Relational World Models
Preprint posted: 2026-11-04 · Original record ↗
- arxiv
- 2611.06653
Explicit source versions (1)
- v1 · 2026-11-04
- Preprint available
Source date: 2026-11-04 · First detected: 2026-11-05 UTC · Historical baseline
Source evidence (2)
- Official evidence ↗ · RSS guid=oai:arXiv.org:2611.06653v1; arxiv:announce_type=new; item pubDate (official metadata)
- Official evidence ↗ · atom:published (official metadata)
Abstract · arxiv
A robot action can satisfy shape and range constraints yet carry the wrong semantics: axes may be permuted, signs inverted, rotations rescaled, or gripper polarity reversed. We study learned state-action-effect relations as offline contracts for detecting these mismatches and ranking their likely fields. Clean held-out transitions calibrate alarms; a fixed inverse vocabulary supports closed-set localization. On ten LIBERO Spatial tasks comprising 432 source episodes, a capacity-controlled comparison at 2.5% empirical clean false-positive rate gives 46.1% recall for a ridge forward residual, 61.0% for a four-component joint Gaussian mixture, and 65.7% for the exact conditional of that same mixture. The joint score retains a 14.9 percentage-point advantage over ridge, while conditioning improves recall by a further 4.7 points. Joint scoring instead has higher localization: 95.8% versus 94.6%. These matched-rate comparisons are retrospective diagnostics, not deployment calibration guarantees. Earlier Object/Goal results broaden tasks within the LIBERO convention; DROID, all three HYDRA tasks, short temporal faults, small offsets, exact repair, and simulated mitigation fail their respective gates, while xArm contributes one task only. The evidence supports a bounded offline diagnostic and a detection-localization tradeoff, rather than joint-density superiority or cross-robot safety.
