DeLTa

DeLTa: Demonstration and Language-Guided

Novel Transparent Object Manipulation

Our goal is to enable robots to execute manipulation tasks on novel transparent objects
from a task instruction, using only a single human demonstration per primitive action.

Human Demonstration

→

Robot Manipulation

Abstract

Despite the prevalence of transparent objects in everyday environments, transparent robotic manipulation research remains limited to short-horizon grasping with poor generalization to novel objects. To address this, we propose DeLTa (Demonstration and Language-Guided Novel Transparent Object Manipulation), a framework that integrates transparent object perception, demonstration-based trajectory retargeting, and vision-language planning for precise long-horizon manipulation from natural language instructions. A key advantage of our method is its single-demonstration approach, which retargets object-centric 6D trajectories to novel transparent objects without category-level priors or additional training. We further present a task planner that refines VLM-generated plans under single-arm eye-in-hand constraints and composes 15 primitive actions derived from 6 human-demonstrated skill trajectories for long-horizon manipulation. Across 122 real-world trials on 6 tasks with 31 objects under varied lighting, occlusion, and tight-space scenarios, DeLTa significantly outperforms existing methods, particularly in long-horizon tasks requiring precise manipulation.

Long-Horizon Task Demonstration

Method

Overview of our DeLTa framework. our framework comprises three key components for precisely manipulating transparent objects in long-horizon tasks from the language instruction. By leveraging vision foundation models, 6D object trajectories are extracted from single-object demonstration videos. These trajectories transfer to novel objects through our last-inch motion planner without requiring additional training.