nalinraut

4 bulan yang laludiscuss

5 bulan yang laludiscuss

MARS - Multi Asset Reconstruction for Simulation Transform 2D images into physics-ready 3D scene assets for robotics training.

Overview

MARS (Multi Asset Reconstruction for Simulation) is a complete pipeline that:

- Detects objects using hybrid vision-language models (Qwen 2.5 VL + GroundingDINO)

- Segments objects from images using SAM (Segment Anything Model)

- Reconstructs full 3D geometry and textures using SAM 3D Objects

- Estimates physics properties (mass, friction, inertia)

- Validates scenes with PyBullet physics simulation

- Exports to multiple formats (USD, MJCF, URDF)

Key Features

- Rich TUI Summary: Pipeline displays a comprehensive summary table at completion showing all detected objects, segmentation results, reconstruction status, physics properties, and validation results

- Prefect Integration: Full workflow orchestration with Prefect, including plain-text logging compatible with Prefect's logging system

- Configurable Models: Support for multiple model variants (Qwen 3B/7B, GroundingDINO tiny/base)

- Intelligent Filtering: NMS-based duplicate removal and configurable area filters with include_background option for large objects

Submissions

Show HN: Inferential – Multi-robot inference scheduling on shared GPUs

Multi-Asset Reconstruction for Simulation

comments