Abstract: To address the challenges that online optimization of operational decision-making and control in complex industrial systems cannot be realized, this talk proposes a unified structure and algorithm for integrated operational optimal decision-making and control, by combining control, optimization, and prediction with AI technology.

It also proposed a parameter self-learning and self-optimizing algorithm for operational optimal decision-making and control integrating systems, by combining mechanism analysis with deep learning, and digital twin with reinforcement learning. An intelligent integrated system for optimal decision-making and control based on end-edge-cloud collaboration is developed, by combining the proposed method and the end-edge-cloud collaboration technology of Industrial Internet.

The system includes the end-optimal decision-making and control system under actual operation, as well as the cloud-edge collaborative parameter self-learning and self-optimizing system for decision-making and control system. The parameter self-learning and self-optimizing system in digital space collaborates with the optimal decision-making and control system under actual operation. Thus, the end-operational optimal decision-making and control system realizes self-learning and self-optimization.

The system has been successfully applied in the energy intensive equipment—fused magnesium furnace and achieved remarkable results in the reduction of carbon emission.

Abstract: Large-scale, dynamic multi-objective optimization problems (LSDMOPs) extend traditional DMOPs into high-dimensional decision spaces, reflecting the growing complexity of real-world dynamic systems. A typical example is the economic control of gas transportation in large-scale pipeline network, which involve hundreds or even thousands of decision variables, representing pipeline node pressures, compressor power settings, and valve operations.

The system is simultaneously affected by various time-varying factors, such as consumer demand, market fluctuation, environmental temperature and humidity. In such contexts, algorithms are required to quickly find optimal solutions after each environment change, maintaining supply-demand balance while minimizing energy consumption.

In practice, there are many more such examples, including dynamic resource allocation of 5G networks in dense urban environments, UAV swarm dispatch in a disaster relief scenario, and large-scale dynamic vehicle routing planning, just to name a few. However, the effectiveness of existing dynamic multi-objective evolutionary algorithms is severely limited for LSDMOPs, due to inadequate training data, predictions in unknown environments, and large-scale dynamic search spaces.

To address these challenges, we propose a diffusion learning-based evolutionary framework, inspired by the intrinsic analogy between iterative evolution of optimization search and stepwise denoising in diffusion learning. Specifically, a new training paradigm is designed to learn the changing patterns of optimal regions in dynamic fitness landscapes.

It achieves this by using populations’ evolutionary trajectories from initial solutions towards Pareto-optimal solutions across historical environments as rich supervised training data. In addition, we introduce a trajectory alignment loss which encourages the stepwise denoising process to conform to the true population evolutionary behaviors in terms of spatial exploration, convergence trends, and boundary adaptation.

The trained model can gradually control denoising direction and intensity using predefined conditions, allowing it to generate optimization paths from random noise toward Pareto-optimal solutions for a new environment.

Along with an adversarial autoencoder-based large-scale dynamic multi-objective evolutionary framework, we will assess how deep generative modeling techniques and large-scale multi-objective evolutionary algorithms can be seamlessly integrated to solve large-scale DMOPs effectively and efficiently.

Experimental results on a typical dynamic multi-objective test suite with problem settings from 10 to 1,000 dimensions demonstrate that the optimization performance of the proposed framework outperforms existing state-of-the-art designs. Especially in large-scale scenarios, the proposed framework is considered superior in terms of solution quality and computational efficiency.

Abstract: In this talk, I will discuss various aspects of embodied intelligence that I have been exploring over the years and present several case studies. These will range from morphological approaches to sensory-motor learning and cognitive reasoning.

While many of these cases showcase successful examples of specialized forms of embodied intelligence, I want to raise an important question: Can we continue to engineer intelligent machines using our current methods, which compartmentalize intelligence into closed, singular forms? Or do we need a new methodology or a fundamental paradigm shift in how we think about embodied intelligence? First, I will examine tactile intelligence in robots.

Throughout our lives, we rely on our sense of touch, a versatile ability that provides awareness of the world and shapes our daily experiences. Significant advancements in whole-body tactile sensing for robots have led to the development of what we call “tactile intelligence,” which extends beyond the traditional capabilities of robotic systems. This advancement allows robots to sense their environment and physically interact with it in ways that resemble human behavior.

I will share several examples of its applications, including whole-body interaction with humanoid robots, human-robot interaction, improved locomotion for humanoid robots, robot navigation among movable objects, locomotion control for bipedal exoskeletons, and robot collaboration. Furthermore, there are still considerable challenges to overcome to fully realize intelligence in robotics. Thus, I will present aspects of robot learning from others.

Learning from human demonstrations is one of the most powerful mechanisms for enhancing capabilities in both humans and robots. Here, I show three methods for human and robot learning from demonstrations: direct mimicry, goal selection, and purposive learning, which I believe will take Physical AI to new heights. I will illustrate aspects of these mechanisms through several real-world examples.

Abstract: Many robotic applications require a robot to operate in an environment with unknown or uncertainty. The robot must rely on sensing and perception to feel its way around, and perception and motion need to be coupled tightly in real time: perception guides motion, while motion enables better perception.

I will discuss our progress in online autonomous perception, planning, learning, and motion of a robot to achieve manipulation tasks such as contact-rich, complex assembly and dense packing, manipulation of model-free deformable linear objects, estimation of physical parameters of objects, and so on. If time allows, I will introduce our work on searching semantic objects in unknown environments and visual coverage of large unknown structures.

Abstract: One of the key challenges in underwater robotics is the difficulty in obtaining accurate location information, either for the robots themselves or for the targets they are tasked with tracking. In this talk I will present several examples from our recent work on estimation and control for underwater robotic systems operating under localization constraints.

I will first discuss a distributed estimation framework for cooperative localization and tracking, in which a team of robots tracks a moving target (e.g., an acoustically tagged fish) using time-difference-of-arrival (TDOA) measurements of signals emitted by the target. I will then introduce a control barrier function-based approach for incorporating observability constraints into motion planning and control, and demonstrate its application to target tracking using range-only measurements.

Finally, I will present the problem of adaptive environmental sampling under localization uncertainty, and discuss how multi-fidelity Gaussian process models can be leveraged to make more effective use of collected measurements for reconstructing spatial environmental fields. Experimental results will be presented throughout the talk to illustrate the proposed approaches.

Abstract: Embodied intelligence is emerging as a key direction in the next stage of artificial intelligence, with broad application potential in industrial manufacturing, autonomous driving, logistics and transportation, home services, healthcare, and elderly care.

Unlike traditional artificial intelligence systems that mainly focus on perception, reasoning, and decision-making in digital environments, embodied intelligence emphasizes the ability of intelligent agents to perceive, understand, interact with, and adapt to the physical world through bodies, sensors, and actions. This talk will briefly review the development trajectory from the recent wave of artificial intelligence to the rise of embodied intelligence.

It will discuss why advances in large models, multimodal perception, robotics, and real-world interaction have created new opportunities for intelligent systems to move from “understanding the world” toward “acting in the world.” The talk will also introduce the core elements of embodied intelligence, including perception, cognition, action, learning, and environmental interaction, as well as several major technical pathways for realizing embodied intelligent systems.

Finally, the talk will highlight representative application scenarios and discuss the future challenges and opportunities of embodied intelligence in both research and industry.

Abstract: For many years, the name of the game in avdanced, human-centric Robotics has been Human-Robot Interaction. In recent years, we have witnessed a further deepening of the relationship between humans and technology. Robotic technologies have been providing definite advances to assist people in need of physical help, including rehabilitation and prosthetics.

Working in fields were humans are placed right at the center of the technology, on the other hand, is helping refocus our robotics research itself. In prosthetics, the goal is to have an artificial limb to move naturally and intelligently enough to perform the task that users intend, without requiring their attention.

By abstracting this idea, a robot of the future can be thought as a physical “prosthesis” of its user, with sensors, actuators, and intelligence enough to interpret and execute the user intention, translating it in a sensible action of which the user remains the owner. In the talk I will present how human-robot integration reaches beyond prosthetics and rehabilitation applications to industrial environments.

For example, exoskeletons and supernumerary limbs for augmenting human possibilities and shared-autonomy robotic avatars, with the robot executing the human’s intended actions and the human perceiving the context of his/her actions and their consequences.