ROMA: LLM System for Real-World Object-Centric Multi-Sensory Active Perception

arXiv:2610.06955 · cs.RO, cs.CV · Submitted 2026-10-03 · Read on arXiv

cs.RO, cs.CV

Submitted: 2026-10-03

Updated: 2026-10-08

Code: https://github.com/GeWu-Lab/ROMA

Project page: https://gewu-lab.github.io/ROMA

Terminology

• Active Perception — 1× in this paper · explained in 2 episodes

Sources

• Qwen3.5-Omni Technical Report
• Digitizing Touch with an Artificial Multimodal Fingertip
• MLA: A Multisensory Language-Action Model for Multimodal Understanding and Forecasting in Robotic Manipulation
• Universal Manipulation Interface: In-The-Wild Robot Teaching Without In-The-Wild Robots
• ActiveVLA: Injecting Active Perception into Vision-Language-Action Models for Precise 3D Robotic Manipulation
• ESI-Bench: Towards Embodied Spatial Intelligence that Closes the Perception-Action Loop
• SaPaVe: Towards Active Perception and Manipulation in Vision-Language-Action Models for Robotics
• Tactile MNIST: Benchmarking Active Tactile Perception
• Multi-Modal Manipulation via Multi-Modal Policy Consensus
• Octopi: Object Property Reasoning with Large Tactile-Language Models
• TacReasoner: A Dynamic Tactile-Language Framework for Interactive Reasoning in Real-World Scenarios
• VitaTouch: Property-Aware Vision-Tactile-Language Model for Robotic Quality Inspection in Manufacturing
• UniTac: A Unified Multimodal Model for Cross-Sensor Tactile Understanding and Generation
• Touch-R1: Reinforcing Touch Reasoning in MLLMs
• Qwen2.5-Omni Technical Report
• Re-Initialization Token Learning for Tool-Augmented Large Language Models
• Qwen3-Omni Technical Report

Related papers

• FMT x: An Efficient and Asymptotically Optimal Extension of the Fast Marching Tree for Dynamic Replanning
• MPCFormer: A physics-informed data-driven approach for explainable socially-aware autonomous driving
• RoboLab: A High-Fidelity Simulation Benchmark for Analysis of Task Generalist Policies
• ARC: A Reasoning Recipe for Robot Foundation Models
• HRDexDB: A 4D Dexterous Grasping Dataset Across Human and Multiple Robot Embodiments
• APT: Action Expert Pretraining Improves Instruction Generalization of Vision-Language-Action Policies