Contact Windows Are a Model-Error Bottleneck, Not a Decision Bottleneck
Event-based and semi-Markov world models rest on a specific claim: on a fixed 50 Hz time axisstance, flight, impact and recovery share one transition function, so prediction errorconcentrates in the contact window and an explicit event time should improve control. We test iton humanoid landing. The error is real: one-step error at touchdown is 4.7x its pre-contactlevel, 2.4x after removing its ordinary growth with horizon. By physical quantity the jump livesentirely in velocity channels (joint velocity 3.2x), while the pose errors the cost actuallyscores show no extra jump (0.67x, 0.79x). On the decision side the answer is no: overwriting themodel with ground truth for three steps cuts failures by 8.1 pp at contact, by 16.1 pp eightsteps later, 20.3 pp at random steps, and a time-matched control recovers only 1.7 pp---mostlyfrom the touchdown frame (78%). The interface should encode the state jump, not the clock: fourcontact-time estimators are all null, while a post-impact reset head gives +0.016 orderingaccuracy and -1.9 pp of failures, with a placebo built from the model's own increment recoveringnone of it. Both effects scale with base-model accuracy.
Authors
- Haihong Qiu
Institutions
- Guangdong University of Technology (CN)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-21
- DOI
- https://doi.org/10.5281/zenodo.24421074
- Primary Topic
- Reinforcement Learning in Robotics
- Type
- preprint