DreamLedger: Execution-Settled Credit Files for World-Model Imagination in Robot Decision Loops

Engineering notes

Engineering notes will be added by the Robot Papers editorial team.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为 VLA、具身智能、人形机器人控制、机器人操作等高价值论文补充中文说明。

Original abstract

Robots are beginning to act on world-model predictions, yet reliability is still expressed through instantaneous, model-internal signals. DreamLedger instead treats reliability as a persistent deployment object: an execution-settled credit file recording how often consumed predictions are borne out, indexed by operating condition, region, and prediction horizon, and consulted before each use. Each consumed prediction is registered as a claim; attributable outcomes are settled against arriving reality at zero labeling cost, an attribution stage excludes measurement-contaminated outcomes, and a settlement-supervised head complements sparse bins. The resulting credit gates consumption: low-credit predictions shorten the dependent horizon or trigger additional observation; every reliance event remains auditable via dependency tickets and replayable logs. We evaluate DreamLedger in three simulated domains (indoor flight, tabletop manipulation, 2D navigation), via mounts on unmodified DreamerV3, TD-MPC2, and V-JEPA 2-AC, and on a real Franka manipulator. Claim failure is dose-monotone in all 12 held-out condition-horizon cells. Credit-gated planning reduces burned imagination (consumed claims that later fail to redeem) by 62% (95% CI 43-81%) versus blind consumption, with equal success and comparable collision rates. At matched risk targets, persistent books cut verification probes from 1.00 to 0.36/episode in manipulation, at success 0.94 versus 0.98; settlement-grounded calibration retains moderate, seed-consistent operating points unlike raw instantaneous gates. The same trust layer operates across decoder-, latent-, and token-space interfaces, including V-JEPA 2-AC settled on real robot frames. On hardware, settlement remains operational under real sensing and contact noise, a deployment failure loop is re-priced online, and all 1,062 registered spends replay from the audit logs.

Links and sources

• Official / arXiv page
• PDF from original source

Comments