Uncertainty-Aware Embodied Artificial Intelligence with Hybrid Reinforcement Learning and Model Predictive Control for Resilient Heterogeneous Multi-Robot Systems

Abstract Embodied artificial intelligence (Embodied AI) has rapidly emerged as a fundamental paradigm for developing autonomous robotic systems capable of perceiving, reasoning, and acting within complex physical environments. Recent advances in foundation models, Vision-Language Models (VLMs), and Vision-Language-Action (VLA) architectures have significantly improved high-level semantic reasoning for robotics; however, their deployment in real-world field environments remains limited by distributional shifts, multimodal uncertainty, dynamic environmental changes, and the absence of safety-aware decision-making mechanisms. These challenges become substantially more pronounced in heterogeneous multi-robot systems operating under partial observability and communication constraints. This paper presents an uncertainty-aware embodied AI framework that integrates multimodal perception, shared latent world modelling, predictive uncertainty estimation, online self-supervised representation adaptation, and hierarchical decision-making through hybrid Reinforcement Learning and Model Predictive Control (RL-MPC). Unlike conventional embodied AI architectures that execute fixed perception and control pipelines, the proposed framework continuously maintains and updates an internal world representation, enabling adaptive reasoning, uncertainty-aware planning, and safe real-time control. Furthermore, the architecture naturally extends to heterogeneous teams of aerial robots, autonomous ground vehicles, and mobile manipulators through distributed latent representations that support cooperative perception, coordinated task allocation, and resilient autonomous behavior. The proposed framework is evaluated in photorealistic NVIDIA Isaac Sim environments containing randomized terrain geometry, vegetation density, illumination changes, dynamic obstacles, sensor degradation, communication uncertainty, and severe out-of-distribution conditions representative of realistic field robotic deployments. Experimental results demonstrate consistent improvements over vision-based navigation, reinforcement learning, model predictive control, and hybrid RL-MPC baselines. The proposed method achieves approximately 21% higher navigation success, reduces collision frequency by more than 35%, exhibits significantly improved robustness under increasing environmental uncertainty, and maintains stable cooperative performance across heterogeneous multi-robot teams. These results demonstrate that uncertainty-aware embodied world models, continual representation learning, hierarchical VLA-guided decision making, and hybrid learning-control architectures provide a scalable foundation for resilient embodied intelligence and represent a promising direction toward next-generation autonomous robotic systems operating safely and reliably in complex real-world environments.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-13
DOI
https://doi.org/10.21203/rs.3.rs-10922191/v1
Primary Topic
Reinforcement Learning in Robotics
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Uncertainty-Aware Embodied Artificial Intelligence with Hybrid Reinforcement Learning and Model Predictive Control for Resilient Heterogeneous Multi-Robot Systems

Babak Rahmani, Fatemeh Afghah
1 links
Research Square
Reinforcement Learning in Robotics
preprint

Uncertainty-Aware Embodied Artificial Intelligence with Hybrid Reinforcement Learning and Model Predictive Control for Resilient Heterogeneous Multi-Robot Systems

Babak Rahmani, Fatemeh Afghah
preprint en

Abstract

Abstract Embodied artificial intelligence (Embodied AI) has rapidly emerged as a fundamental paradigm for developing autonomous robotic systems capable of perceiving, reasoning, and acting within complex physical environments. Recent advances in foundation models, Vision-Language Models (VLMs), and Vision-Language-Action (VLA) architectures have significantly improved high-level semantic reasoning for robotics; however, their deployment in real-world field environments remains limited by distributional shifts, multimodal uncertainty, dynamic environmental changes, and the absence of safety-aware decision-making mechanisms. These challenges become substantially more pronounced in heterogeneous multi-robot systems operating under partial observability and communication constraints. This paper presents an uncertainty-aware embodied AI framework that integrates multimodal perception, shared latent world modelling, predictive uncertainty estimation, online self-supervised representation adaptation, and hierarchical decision-making through hybrid Reinforcement Learning and Model Predictive Control (RL-MPC). Unlike conventional embodied AI architectures that execute fixed perception and control pipelines, the proposed framework continuously maintains and updates an internal world representation, enabling adaptive reasoning, uncertainty-aware planning, and safe real-time control. Furthermore, the architecture naturally extends to heterogeneous teams of aerial robots, autonomous ground vehicles, and mobile manipulators through distributed latent representations that support cooperative perception, coordinated task allocation, and resilient autonomous behavior. The proposed framework is evaluated in photorealistic NVIDIA Isaac Sim environments containing randomized terrain geometry, vegetation density, illumination changes, dynamic obstacles, sensor degradation, communication uncertainty, and severe out-of-distribution conditions representative of realistic field robotic deployments. Experimental results demonstrate consistent improvements over vision-based navigation, reinforcement learning, model predictive control, and hybrid RL-MPC baselines. The proposed method achieves approximately 21% higher navigation success, reduces collision frequency by more than 35%, exhibits significantly improved robustness under increasing environmental uncertainty, and maintains stable cooperative performance across heterogeneous multi-robot teams. These results demonstrate that uncertainty-aware embodied world models, continual representation learning, hierarchical VLA-guided decision making, and hybrid learning-control architectures provide a scalable foundation for resilient embodied intelligence and represent a promising direction toward next-generation autonomous robotic systems operating safely and reliably in complex real-world environments.

Research Square
IEEE Foundation (US)
Reinforcement Learning in Robotics
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon