Image-based robot navigation with task achievability.

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Image-based robot action planning leverages deep learning advancements.
• Current methods rely on estimating optimal cost-minimizing paths using parametric models, often deep neural networks.
• These models require extensive labeled data, which is challenging to collect in real-world robotic tasks.

Purpose of the Study:

• To address the inaccuracy of cost estimation models trained with robot-autonomously collected data.
• To introduce and evaluate a novel metric, "task achievability" (TA), for robot navigation.
• To demonstrate TA's effectiveness in improving robot navigation success rates.

Main Methods:

• Investigated the performance degradation of parametric cost estimators trained on robot-collected data.
• Proposed "task achievability" (TA) as a metric defined by the probability of reaching a goal within a time limit.
• Compared TA-based training with traditional optimal cost estimator training, utilizing both optimal and non-optimal trajectories.

Main Results:

• Models trained with robot-collected data showed inaccurate cost estimations, leading to navigation failures.
• TA-based training demonstrated stable estimation and superior performance compared to optimal cost estimators.
• Robot navigation experiments in a living room environment confirmed TA's success in reaching target positions.

Conclusions:

• Traditional cost estimation methods are unreliable when trained on robot-collected data.
• Task achievability (TA) offers a robust alternative metric for robot action planning.
• TA enables successful robot navigation even in challenging scenarios where conventional methods fail.