Go to ICLR 2026 Conference homepageBridging the Perception Gap: Probe-Guided Data Optimization Framework for Robotic Imitation Learning
Keywords: Imitation Learning, Manipulation, Proxy Model
TL;DR: A new method that optimizes the data collection process using a proxy model. This approach bridges the perceptual gap between humans and robots, improving data collection efficiency, model robustness, and success rates.
Abstract: Imitation learning allows robots to acquire complex manipulation skills from human demonstrations. However, traditional data collection methods often haven't account for the "perceptual gap" between humans and robots, which leads to models that don't perform as expected. To solve this, we introduce Policy-Intent Probe (PIP). This method trains a proxy model with a small amount of demonstration data, then quantifies the model's perceptual capabilities based on its policy distribution. Based on the model's feedback, we divide the task's operational space into a Robust Operation Zone (ROZ) and a Non-Robust Operation Zone (NROZ). By standardizing the trajectories in the NROZ and waiting until the robot enters the ROZ to perform precise operations, we have optimized the data collection trajectories. Furthermore, aided by PIP, we can correctly carry out subtask segmentation. We can supplement data collection based on subtask complexity, which enhances the model's generalization and robustness. By cleaning only the subtask data containing anomalous trajectories or failure, we minimize data loss. Based on an empirical evaluation on three real-world tasks, we proved that perceptual capabilities can affect a task's success rate and that arbitrary subtask decomposition can lead to negative consequences. Our model-in-the-loop data optimization framework can significantly boost the success rate of long-horizon precision manipulation tasks, enhance model robustness, and increase data collection efficiency.
Primary Area: applications to robotics, autonomy, planning
Submission Number: 24732
Loading