Demonstration-conditioned policies let a robot infer what to do from example trajectories, but they struggle with long-horizon tasks. The arXiv paper "PACE: Stage-Consistent Long-Horizon Robot Manipulation via Progress-Aligned Context for Execution" identifies a specific failure mode: when visually similar states recur at different stages of a task, the policy can lose track of which stage it is in.
To address this, the authors propose PACE, which stands for Progress-Aligned Context for Execution. The method is designed to keep the policy's context aligned with the actual progress of the task, so that execution remains consistent with the current stage rather than being misled by recurring visual features.
The abstract is truncated, so full method details and experimental results are not available in this source. This is a single-source digest, so there are no other sources to compare or contrast.