Journal of Control and Decision

Bridging Deep Visual Representations and Geometric Control for Embodied Intelligence

Theme:
The rapid advancement of foundation models and Vision-Language Models (VLMs) has brought unprecedented perceptual capabilities to autonomous systems. These deep networks excel at compressing complex, high-dimensional sensory inputs into rich, low-dimensional latent embeddings that inherently capture semantic structures, viewpoint equivariance, and object invariance.

However, the deployment of these data-driven representations in physical, safety-critical embodied artificial intelligence (AI) systems remains fundamentally challenged by a theoretical gap. Classical nonlinear control and estimation theories rely on explicit physical state spaces, rigorous geometric structures, and analytical models, which are often incompatible with the black-box nature of deep visual embeddings.

This Special Issue aims to bridge this gap. We seek to explore how the unstructured, high-dimensional outputs of modern perception models can be treated as novel "state manifolds" to synthesize control laws, design observers, and guarantee safety. By unifying deep visual representation learning with geometric control theory, this issue will catalyze the transition toward robust, theoretically grounded embodied intelligence.

We invite original research articles, comprehensive reviews, and perspective papers that explore the intersection of representation learning, state estimation, and nonlinear control.

Topics of interest include, but are not limited to:

• Control and Planning in Latent Spaces: Methodologies for designing feedback controllers, visual servoing, and motion planning algorithms directly on high-dimensional visual embedding manifolds.
• Safety and Robustness with Implicit Representations: Formulations of safety bounds, invariant sets, and obstacle avoidance strategies (e.g., potential fields, barrier functions) operating on data-driven visual features.
• Data-Driven Dynamics and Linearization: Techniques for modeling system dynamics, extracting invariant subspaces, or achieving global/local linearization from complex multimodal observation streams.
• State Estimation and Observers on Data Manifolds: Novel filtering and observer design bridging Euclidean embedding spaces and intrinsic, low-dimensional physical manifolds without explicit pose estimation.
• Geometric Properties of Visual Embeddings: Theoretical analyses of equivariance, invariance, and topological structures within deep latent spaces, and their implications for robotic control.
• Sim-to-Real and Domain Generalization: Utilizing the robust nature of foundation models to improve the transferability of control policies across varying visual and physical domains.

This Special Issue targets a multidisciplinary audience, bridging researchers from control theory, robotics, computer vision, and machine learning. It is particularly relevant for scholars investigating theoretical guarantees for learning-based control and those developing next-generation perception-action loops for autonomous vehicles, unmanned aerial vehicles, and dexterous manipulators.

By moving beyond ad-hoc end-to-end learning and traditional explicit state estimation, this issue will establish a foundational framework for "perception-aware control." It will provide the community with both theoretical insights and the practical methodologies needed to deploy intelligent systems in unstructured, dynamic, and unpredictable real-world environments.

Description of the Publicity and Promotion Plan

The Special Issue will be intensively promoted through multiple international channels to ensure broad visibility and attract high-quality submissions from the control, robotics, and artificial intelligence communities.

The Guest Editors will disseminate the Call for Papers through their professional networks, editorial board memberships, IEEE Technical Committees, and international research collaborations. The Special Issue will also be promoted through relevant conferences, workshops, and special sessions, including ARCAI 2026 (2026 The 3^rd International Conference on Advanced Robotics, Control and Artificial Intelligence).

We will include the following publicity activities:

• Distribution of the Call for Papers through academic mailing lists and professional societies;
• Promotion via conference websites, newsletters, and social media platforms (e.g., LinkedIn, WeChat, ResearchGate);
• Direct invitations to leading researchers and emerging scholars in relevant fields;
• Promotion through partner universities, research institutes, and industry collaborators;
• Recommendations for the further extension of outstanding papers presented at ARCAI 2026 and related international conferences for journal submission.

The above publicity activities are expected to attract a diverse pool of high-quality submissions from researchers worldwide and ensure the scientific impact and visibility of this Special Issue.