Multimodal Agent MemorySubmit a paper

DeLTa

DeLTa: Demonstration and Language-Guided Novel Transparent Object Manipulation

Taeyeop Lee, Gyuree Kang, Bowen Wen, Youngho Kim, Seunghyeok Back, In So Kweon, et al. · 2025-10-07

Read the paperPDF

Management
T1 Static Memory
Memory is built once and reused without change during deployment.
Primary modality
Embodied & 3D
Also: Image, depth, text
Representation
Source- and modality-specific

How experience becomes memory

Each path is one memory component, traced from what is stored to how it is found again.

  1. database of fixed object-centric 6-DoF pick, place and pour demonstration trajectories

    1. Represented asModality-specific encodings & geometric stateSource- and modality-specific
    2. Stored inFlat sequential records
    3. Retrieved byPlanner / tool-assisted retrieval
    4. Source evidenceSource not retained
    The system extracts one object-pose trajectory for each pick, place and pour skill, transforms it into an action-specific reference frame, subsamples and smooths it, and explicitly stores the processed trajectories in a trajectory database.Full paper: Sec. III-A Parsing Human Video Demonstration · PDF p. 4 · Figure 2; Trajectory Database paragraph
    During robot execution, the action manager selects the stored trajectory-based skill requested by the task plan, retrieves and retargets its object-frame waypoints to the current object, and follows the aligned demonstration trajectory through the last-inch manipulation stage.Full paper: Sec. III-C Demonstration-Guided Robot Action Execution · PDF p. 5 · Figure 5; Eq. 2-5

Related systems