Drift is a Sampling Error: SNR-Aware Power Distributions for Long-Horizon Robotic Planning

2026-05-10

AI Summary

This work addresses the instruction drift problem in vision-language-action models during long-horizon robotic tasks, which arises from locally greedy sampling strategies. To mitigate this issue, the authors propose Context-Aware Power Sampling (CAPS), a training-free, inference-time adaptive computation framework. CAPS enhances global trajectory likelihood using a power distribution and incorporates a signal-to-noise ratio (SNR)-based metacognitive mechanism that dynamically triggers MCMC search upon detecting drift risk, enabling seamless transitions from fast intuitive decisions to slow deliberative planning. By formally modeling instruction drift as a systematic sampling bias and integrating power-law reweighting with SNR-aware monitoring, CAPS significantly improves robustness in extended tasks. Experiments demonstrate consistent and substantial performance gains over strong baselines—including OpenVLA and TACO—across RoboTwin, Simpler-WindowX, and Libero-long benchmarks.

Abstract

Despite rapid progress in Vision-Language-Action (VLA) models for robotic control, instruction drift remains a persistent failure mode in long-horizon tasks. This paper reconceptualizes this phenomenon, positing that instruction drift is fundamentally a systematic sampling error: local greedy sampling is prone to collapsing into "Negative Pivotal Windows"--irreversible local optima with high local probability that sever global success pathways. To address this, we propose Context-Aware Power Sampling (CAPS), a training-free inference-time computation framework. CAPS leverages power distributions to sharpen global trajectory probabilities, enabling lookahead search over the model's conditional generative trajectory distribution. Furthermore, we introduce a metacognitive control mechanism based on Signal-to-Noise Ratio (SNR). This mechanism triggers adaptive MCMC search solely when drift risk is detected, enabling a dynamic transition from "intuitive fast thinking" to "rational slow search." Experiments on RoboTwin, Simpler-WindowX, and Libero-long benchmarks show that CAPS achieves substantial improvements over strong baselines, including OpenVLA and TACO, without parameter updates. These results support the effectiveness of adaptive inference-time computation for improving long-horizon robustness in embodied control.

Problem

• instruction drift
• sampling error
• long-horizon robotic planning
• local optima
• embodied control

Innovation

• instruction drift
• power distributions
• SNR-aware control
• adaptive inference
• long-horizon planning