Robotics

PRISM: Personalized Robotic Dataset Generation via Image-based Scene and Motion Synthesis

Dogyu Ko, Haneul Kim, Chanyoung Yeo, Dowoon Lee, Taeho Park, Hyoseok Hwang

Featured July 13, 2026

This analysis was generated by SciGrove. Upload your own PDFs or enter a DOI — and get the same AI breakdown on any paper.

Get started

AI-generated analysis — This is SciGrove's AI interpretation of the paper, not peer-reviewed content. Always refer to the original paper.

Simply

PRISM creates custom robot training data from just one picture and a command, building diverse virtual scenes that look like the real world but have different objects, helping robots learn tasks better and adapt to new places.

In depth

PRISM addresses the challenge of generating personalized robotic datasets for user-specific environments from a single image and natural language instruction. It achieves this by constructing digital cousin scenes that maintain semantic and geometric alignment with the target while introducing instance-level diversity. Furthermore, it synthesizes executable demonstrations using a motion-aware grasp selection strategy and efficiently expands datasets via trajectory-preserving visual randomization.

Key Takeaways

1
PRISM generates personalized robotic datasets from a single image and natural language, bridging the gap between simulation scalability and real-world alignment.
2
It introduces digital cousin scenes that preserve target environment structure while offering instance-level diversity, crucial for generalization.
3
The pipeline incorporates motion-aware grasp selection for natural demonstrations and trajectory-preserving visual randomization for efficient, appearance-invariant policy learning.

Conceptual Flow

HIGH LEVEL

Methodology (The 'Logic')

The system takes a picture and a command, then builds many similar virtual scenes with different objects, and creates robot actions for them.

Real World Picture

Task Command

Build Virtual Worlds & Actions

Robot Training Data

Results (The 'Impact')

Robots trained with this new data perform much better in both familiar and new environments compared to older methods.

Old Training Data

New Training Data

Train Robot Brains

Better Robot Performance