Problem
Research questions and friction points this paper is trying to address.
reliability
world-model predictions
robot decision loops
instantaneous signals
DreamLedger: Execution-Settled Credit Files for World-Model Imagination in Robot Decision Loops
🤖 AI Summary
DreamLedger通过建立执行结算信用文件来提高机器人决策循环中世界模型预测的可靠性，减少错误预测消耗，已在模拟和真实环境中验证有效。
📝 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.
Innovation
Methods, ideas, or system contributions that make the work stand out.
execution-settled credit file
reliability
world-model predictions
robot decision loops
settlement-supervised head
🔎 Similar Papers
No similar papers found.
