Automation Science at a Crossroads: Charting the Path Toward Truly Intelligent Systems
In a quiet laboratory nestled within Southeast University’s School of Automation, a team of researchers watches intently as a fleet of micro-drones—each no larger than a hand—executes a synchronized dance above a simulated disaster zone. One drone drops a sensor pod; another maps terrain in real time; a third relays encrypted data to a ground station where a human operator makes high-level decisions. There’s no central controller issuing commands. Instead, each unit negotiates its role on the fly—adapting, learning, and compensating for the sudden failure of a neighboring node. This isn’t science fiction. It’s the front line of a quiet but profound revolution: the convergence of automation science and artificial intelligence into autonomous intelligent systems—machines that don’t just react, but reason, collaborate, and evolve.
For decades, automation has quietly powered modern civilization—from the precision timing of traffic lights to the millisecond-level control loops stabilizing jet engines. But as the world barrels toward hyperconnectivity, climate urgency, aging populations, and geopolitical uncertainty, the old paradigms are cracking under strain. What happens when a factory’s assembly line must reconfigure overnight to produce ventilators? When a deep-sea mining robot encounters an unknown hydrothermal vent system? When a Mars rover loses contact with Earth for 22 minutes—and must decide whether to descend a cliff?
The answers, according to a landmark analysis published in Acta Automatica Sinica, hinge not on incremental upgrades, but on reimagining automation itself—from a discipline focused on control to one centered on co-evolution: between humans and machines, physical and digital worlds, certainty and ambiguity.
Led by Chang-Yin Sun of Southeast University and a coalition of China’s top automation scientists—from the National Natural Science Foundation to the Shenyang Institute of Automation—the team delivers a stark yet optimistic diagnosis: automation stands at its most pivotal juncture since the advent of digital control. The challenges are systemic. The opportunities, transformative.
At the heart of this inflection point lies a fundamental mismatch. Classical control theory—elegant, mathematically rigorous, and astonishingly effective for closed-loop systems with known dynamics—was built for predictability. Think cruise control on a highway: measure speed, compare to setpoint, adjust throttle. Repeat. But modern operational environments are anything but predictable. They are open, non-stationary, and multi-stakeholder.
A wind farm doesn’t just face shifting gusts; it contends with grid fluctuations, wildlife migration alerts, and real-time electricity pricing signals. A surgical robot doesn’t merely track anatomical landmarks; it must interpret a surgeon’s subtle hand tremor as either fatigue or intentional micro-movement—and respond accordingly. In such contexts, “model-based” control falters. The model—that idealized mathematical twin of reality—becomes outdated faster than it can be updated.
This isn’t failure; it’s evolution. As Sun and colleagues argue, the field is undergoing a conceptual pivot—from model-driven to data- and intelligence-driven paradigms. But crucially, this isn’t about slapping deep learning onto legacy controllers. It’s about rebuilding the theoretical scaffolding from the ground up.
Consider data-driven control. Rather than assuming a fixed plant model, this emerging framework treats data not as secondary evidence, but as primary ontology. Controllers learn not equations, but invariant relationships—temporal, causal, functional—directly from sensor streams, even as systems degrade or environments shift. It’s akin to how a seasoned pilot doesn’t recalculate aerodynamics mid-turbulence; they feel the aircraft’s behavior and respond intuitively, having internalized thousands of prior flight patterns.
Yet raw data isn’t enough. Noise, sparsity, and adversarial manipulation plague real-world sensing. Hence the urgent need for robust, interpretable learning. The team highlights deep reinforcement learning (DRL) not as a black-box oracle, but as a collaborative reasoning engine—especially when embedded in multi-agent settings. Here, the real breakthrough isn’t individual performance; it’s emergent coordination. Picture a swarm of autonomous cargo ships navigating the South China Sea: each vessel optimizes its own fuel use, but collectively, they minimize regional emissions and collision risk—without a central dispatcher, through local negotiation and shared value estimation. This “cooperative DRL” represents a new class of distributed intelligence, where system-level resilience arises from decentralized cognition.
Still, even the smartest algorithm falters without embodiment. Which brings us to the second frontier: high-performance operational robots.
Forget the humanoid butlers of 1950s sci-fi. Today’s most consequential robots are task-specific virtuosos—designed not to mimic humans, but to augment and extend human capability in extreme or inaccessible domains. China’s push here is both strategic and pragmatic: from deep-ocean mineral extraction to elder-care assistance, the nation faces demographic and resource pressures that demand robotic solutions.
But progress isn’t linear. While Chinese firms now lead in applied domains—facial recognition accuracy, drone delivery logistics, industrial arm deployment—the paper candidly acknowledges structural gaps. Core components remain imports: high-torque harmonic drives from Japan, force-torque sensors from Germany, real-time operating kernels from the U.S. Worse, the theoretical foundation lags. How do you design a legged robot that can leap across fractured ice and delicately retrieve a biological sample—switching gait, compliance, and intent mid-motion? That demands unified frameworks linking biomechanics, material science, and real-time decision theory.
The authors spotlight bio-inspired robotics as a promising bridge. Not mere aesthetic mimicry—like Boston Dynamics’ parkour-capable machines—but deep functional emulation. Take the cuttlefish: its skin dynamically modulates reflectance and texture for camouflage; its hydrostatic muscles allow silent, omnidirectional propulsion. Translating such principles into soft, sensor-rich actuators could yield underwater drones that evade sonar and perform close-proximity inspection without disturbing sediment.
Equally vital is human-robot symbiosis. The future isn’t human or robot—it’s hybrid. In trauma surgery, a robotic arm steadies a scalpel while interpreting neurophysiological signals to anticipate the surgeon’s next move. In disaster response, a first responder wears an exoskeleton that amplifies strength and shares situational awareness with UAV scouts overhead—each enriching the other’s perception. This “bidirectional cognition” requires new interfaces: not buttons and joysticks, but semantic intent mapping, where machines infer goals from gaze, gesture, even EEG micro-patterns, and convey system state through haptic feedback or augmented reality overlays.
Crucially, such systems must be trustworthy. Not just reliable, but explainable and ethically bounded. If a care robot decides to restrain an agitated dementia patient, can it articulate why—weighing fall risk against dignity, referencing prior clinician directives? The paper calls for embedded safety ontologies: formal rules that govern behavior not as rigid constraints, but as negotiable principles, updated through interaction.
If intelligent control and embodied robotics form the “muscle and brain” of next-gen automation, then cyber-physical systems (CPS) are its nervous system—the seamless fusion of computation, networking, and physical process. Yet here, too, legacy assumptions falter.
Early CPS visions imagined perfectly synchronized sensor networks feeding flawless models. Reality is messier: packet drops in 5G tunnels, sensor drift in Arctic cold, adversarial spoofing of GPS signals. The result? Cascading failures. A smart grid’s frequency stabilizer misreads load due to corrupted phasor data; it overcorrects; neighboring nodes misinterpret the correction as demand surge; blackouts propagate.
Sun’s team reframes resilience not as redundancy, but as adaptive elasticity. Three pillars emerge:
First, edge-intelligent control. Instead of funneling all data to a cloud AI for decisions, embed lightweight learning agents at the edge—on PLCs, drone flight controllers, even individual machine tools. These agents detect anomalies locally, negotiate resource sharing (e.g., bandwidth, compute), and execute fallback protocols before failures escalate. Think of it as “immune response” at the component level.
Second, security-by-design integration. Traditional IT security—firewalls, encryption—fails when the physical actuator is the attack vector. If a hacker subtly biases a wind turbine’s pitch angle by 0.5 degrees, fatigue cracks accumulate invisibly until catastrophic failure. The paper advocates co-design: embedding anomaly detection not just in network packets, but in control loop residuals—the tiny discrepancies between expected and actual system behavior that betray manipulation.
Third—and perhaps most revolutionary—digital twin co-evolution. Today’s digital twins are static snapshots. Tomorrow’s must be lifelong learners. A twin of a high-speed train doesn’t just mirror its current state; it simulates counterfactual scenarios: “What if bearing X fails at 300 km/h in -30°C?” It then prescribes maintenance and updates its own physics models based on real-world sensor feedback, closing a continuous learning loop between virtual and physical.
This vision is already taking shape. In Shenyang, researchers at the State Key Laboratory of Robotics are testing twins for industrial forging presses—predicting die wear not from scheduled hours, but from real-time vibration spectra and thermal imaging, extending tool life by 22%. In Tianjin, teams are building “twin hospitals,” where patient flow, staff allocation, and equipment utilization are optimized in simulation before deployment, reducing ER wait times by over 35%.
No discussion of modern automation is complete without confronting navigation and control in the world’s most hostile frontiers: deep space, deep sea, and dense urban canyons.
GPS, the workhorse of modern positioning, fails spectacularly underground, underwater, or amid skyscrapers. Yet China’s ambitions—from lunar south pole bases to Arctic shipping lanes—demand ubiquitous, resilient localization. The answer lies in multi-source fusion, but not as simple sensor averaging.
The paper champions semantic navigation: where systems don’t just triangulate position, but understand context. A Mars rover identifies rock strata not as pixels, but as geological markers—comparing them to orbital maps to self-correct drift. An autonomous submarine uses bathymetric contours not just for “where am I,” but “what is this feature?”—distinguishing a seamount from a shipwreck to avoid collision or trigger sampling.
Even more radical is quantum-aided navigation. While still nascent, quantum accelerometers and gyroscopes—immune to electromagnetic interference and offering drift rates orders of magnitude lower than MEMS—could enable submarines to navigate for months without surfacing, or drones to operate in GPS-denied combat zones. China’s investments here are substantial, with prototypes already tested in laboratory vacuum chambers.
Meanwhile, in cities, the challenge shifts from precision to scalability. Tracking one autonomous taxi is feasible; coordinating 10,000 across Beijing requires rethinking infrastructure itself. The team proposes V2X-native traffic systems: where intersections don’t rely on traffic lights, but on real-time negotiation between vehicles and infrastructure—prioritizing ambulances, smoothing platoons, dynamically reallocating lanes. Early trials in Suzhou show potential to reduce congestion by 40%—not by adding roads, but by intelligently sharing existing ones.
Underpinning all this is the sobering reality of major equipment automation. Jet engines, nuclear reactors, high-speed rail networks—these “national arteries” cannot afford trial-and-error AI. Here, automation’s role shifts from optimization to assurance.
The key insight? Intelligence must serve safety, not vice versa. For instance, data-driven fault diagnosis doesn’t replace physics-based models; it augments them. By correlating subtle vibration harmonics with historical failure modes, systems can flag incipient cracks weeks before they’re visible to inspectors. Similarly, self-healing control reconfigures redundancy on-the-fly: if a thruster fails mid-orbit, the system doesn’t just switch to backup—it recalculates optimal thrust allocation across remaining units, preserving mission objectives.
Critically, such systems require provenance-aware AI. Every decision—especially in safety-critical contexts—must be auditable. Not just “what” the system did, but “why,” based on which data, at what confidence level. This demands new standards for explainable machine learning in control loops, where saliency maps aren’t for researchers, but for certification engineers.
Finally, the paper confronts the elephant in the room: AI-driven automation remains hamstrung by brittleness. AlphaGo’s brilliance on the Go board doesn’t translate to handling a dropped wrench on a factory floor. Why? Because real-world tasks are open-ended, partially observable, and socially embedded.
The path forward, the authors argue, lies in hybrid-augmented intelligence—a term they define not as human + AI, but as complementary cognition. Machines excel at data crunching, pattern spotting, and tireless execution. Humans bring abstraction, ethics, and creative leaps. The magic happens in the handoff: when a robot flags an anomaly it can’t classify, it doesn’t halt—it queries a human expert, learns from the response, and updates its knowledge base. Over time, the boundary of “known unknowns” shrinks.
This requires rethinking development itself. Instead of monolithic AI deployments, the field must embrace modular, upgradable intelligence—like app ecosystems, where perception, planning, and control modules can be swapped as new algorithms emerge. Open benchmarks, shared datasets (with privacy safeguards), and standardized hardware interfaces become as vital as the algorithms themselves.
None of this happens in a vacuum. The paper concludes with a call for ecosystem-level investment: nurturing talent not just in coding, but in systems thinking; building open testbeds for swarm robotics or smart grids; fostering international collaboration even amid competition. Because the challenges—climate adaptation, pandemic resilience, sustainable manufacturing—transcend borders.
Automation, once the silent enabler of the 20th century, is poised to become the architect of the 21st. Not by replacing humans, but by amplifying our capacity to understand, adapt, and care for an increasingly complex world. The drones in that Southeast University lab? They’re not just practicing rescue ops. They’re rehearsing a new covenant: between silicon and synapse, between prediction and presence, between control and co-creation.
The future won’t be automated.
It will be co-intelligent.
Chang-Yin Sun¹, Guo-Zheng Wu², Zhi-Heng Wang², Yang Cong²,³, Chao-Xu Mu⁴, Wei He⁵
¹ School of Automation, Southeast University, Nanjing 210096
² Department of Information Sciences, National Natural Science Foundation of China, Beijing 100085
³ State Key Laboratory of Robotics, Shenyang Institute of Automation, Chinese Academy of Sciences, Shenyang 110016
⁴ School of Electrical and Information Engineering, Tianjin University, Tianjin 300072
⁵ School of Automation and Electrical Engineering, University of Science and Technology Beijing, Beijing 100083
Acta Automatica Sinica, 2021, 47(2): 464–474
DOI: 10.16383/j.aas.c200904