Keskeiset käsitteet
Long-horizon Visual Action (LAVA) enhances robotic food acquisition with a hierarchical policy framework.
Tiivistelmä
I. Abstract:
Robotic Assisted Feeding (RAF) aims to restore independence in feeding for individuals with mobility impairments.
Existing RAF methods focus on solid foods, leaving a gap in strategies for semi-solid and deformable foods.
LAVA introduces a hierarchical policy framework for long-horizon food acquisition.
II. Introduction:
RAF addresses challenges faced by individuals with limited mobility in feeding themselves.
Bite acquisition is crucial for robotic arms to transfer food to the mouth efficiently.
III. Proposed Approach:
LAVA employs high-level, mid-level, and low-level policies for food acquisition.
The framework divides tasks into decision-making, action refinement, and execution levels.
IV. Experiments:
LAVA's networks achieve high accuracy in selecting primitives and predicting bite targets.
Comparison with baselines shows LAVA's superior performance across various food types.
V. Results:
LAVA adapts well to handling liquids prone to spillage and deformable foods like tofu without breakage.
Zero-shot generalization demonstrates robust performance across different food configurations.
VI. Conclusion & Future Work:
LAVA's hierarchical policy framework improves efficiency and adaptability in robotic-assisted feeding.
Limitations exist in handling thin or irregular foods, suggesting future work on diverse food types.
Tilastot
LAVAは46のボウルで89±4%の成功率を達成しました。
ScoopNetは5316枚の画像から高レベルプリミティブを正確に選択します。
DepthNetは175インスタンスで85.7%の精度でスプーンの深さを決定します。