paper_rob__rdd

What problem does it solve?

Decomposes long-horizon robot demonstrations into actionable sub-tasks by retrieving similar segments from a prior database and optimizing partitioning, enabling planning without training.

Core Features & Use Cases

• Embeds frames with pre-trained visual encoders (LIV, CLIP, R3M) to form robust frame representations.
• Retrieves similar sub-task segments from a prior database via approximate nearest-neighbor search.
• Decomposes demonstrations with dynamic programming to maximize retrieval similarity while respecting length constraints.
• Training-free approach: does not require end-to-end model training and supports multiple encoders and multi-view setups.
• Suitable for robotics planning benchmarks, task analysis, and human-robot collaboration scenarios.

Quick Start

Embed frames with a chosen encoder, load the nearest-neighbor database, and run the DP-based partitioning to obtain sub-task segments.