Chongqing University Team Cuts Battery-Swap Robot Energy Use by 11% Through Integrated Design Breakthrough
In an era where electric mobility is rapidly shifting from niche adoption to mass-market inevitability, the bottleneck is no longer just the vehicle itself—it’s the supporting infrastructure. While charging stations crowd highways and urban corners, a quieter but equally vital revolution is unfolding in the world of battery swapping: the rise of robotic assist systems engineered for speed, precision, and, crucially, energy efficiency. At the forefront of this evolution, a research team from Chongqing University has unveiled a novel optimization strategy that slashes energy consumption in battery-swap robots by over 10%—without compromising performance or safety. Their breakthrough? A radical rethinking of how mechanical design and motion control should be co-optimized, not treated in isolation.
The implications extend far beyond academic interest. As electric fleet operators—especially in logistics, public transit, and ride-hailing—push for sub-five-minute battery swaps to rival gasoline refueling, robotic arms must cycle thousands of times per month. Each joule saved per cycle compounds into massive reductions in operational cost and grid demand. Yet most current systems still rely on sequential design: engineers first fix the kinematic layout, then tune the controller. This decoupled approach, while practical, leaves performance on the table. The Chongqing team’s work demonstrates that true efficiency emerges only when structure and control speak the same language—literally, in the syntax of multi-objective optimization.
At the heart of their innovation lies the lifting module of a battery-swap robot—a deceptively simple subsystem that hoists an 80-kilogram EV battery pack through a vertical stroke of 220 millimeters in under 10 seconds. To the untrained eye, it resembles a compact scissor-lift or a linear actuator with linkage arms. But beneath that mechanical elegance resides a tightly coupled dynamical system where every millimeter of link length, every turn of gear ratio, and every gain in the control algorithm reshapes the energy landscape.
The team, led by Associate Professor Lin Lihong, began by mapping the full physical chain: a 400-watt permanent-magnet AC servo motor drives a planetary-style reduction gear (initially set at 1.2:1), which rotates a 12-millimeter-diameter ball screw with a 10-millimeter lead. The screw’s linear motion feeds a planar four-bar linkage—specifically, a slider-crank variant—comprising three articulated rods (labeled l₁, l₂, and l₃ in their model). The final rod connects to a horizontal slot in the battery tray, converting lateral slider movement into vertical lift via constrained kinematics. This architecture offers compactness and rigidity—ideal for garage-floor installations—but introduces nonlinear coupling between motor torque, load inertia, and gravitational resistance.
Early prototypes performed adequately. But “adequate” isn’t enough when scaling to hundreds of swap stations across a city. Initial tests revealed a sobering truth: over 15% of the robot’s total energy budget was spent just raising and lowering the empty or loaded tray—energy that didn’t contribute to battery exchange, safety, or user experience. Worse, position overshoot and settling oscillations prolonged cycle time, forcing conservative scheduling and reducing throughput.
The team refused to accept the conventional trade-off: you can have speed or efficiency, but not both. Instead, they posed a bolder question: What if the robot’s skeleton and nervous system were designed together from day one?
To answer it, they developed a unified framework anchored in two pillars: physics-based modeling and intelligent multi-objective search.
First, dynamics. Rather than treat the mechanism as a black box, they derived its full Lagrangian equations of motion—accounting for every rotating gear’s inertia, every sliding mass’s kinetic energy, and the changing gravitational potential as the tray ascends. Crucially, they included the motor’s electromechanical behavior: back-EMF, torque ripple, and—in a nod to real-world losses—viscous and Coulomb friction at joints and bearings. This yielded a high-fidelity energy consumption model: total input energy equals the time integral of motor input power, where power is the product of commanded torque and angular velocity, adjusted for efficiency losses in transmission and electronics.
Second, control. They chose sliding-mode control—not for its simplicity, but for its robustness. Unlike PID or linear-quadratic regulators, sliding-mode thrives in the presence of parameter uncertainty and external disturbances (say, battery pack mass variation or temperature-induced viscosity changes in grease). Their formulation defined a sliding surface as a linear combination of position error and its derivative, then enforced convergence using an exponential reaching law—a strategy that accelerates initial response while damping chatter near the target. To further suppress high-frequency oscillation (a notorious side effect of pure sign-function switching), they swapped the hard sgn() for a smooth saturation function, preserving robustness while easing mechanical stress.
But here’s where the project diverged from prior art: instead of tuning the sliding-mode gains (ε, λ, k) after the mechanism was built, they treated them as design variables on equal footing with physical parameters—gear ratio i, screw lead Pₕ, and link lengths l₁–l₃. All eight parameters formed a joint decision vector, optimized not for a single goal, but two competing objectives: minimize total lift energy (E) and minimize steady-state angular position error (eₛₛ)—the latter serving as a proxy for tray positioning repeatability and user confidence.
This multi-objective stance was critical. Reducing error alone could mean aggressive control—high gains, fast corrections—leading to torque spikes and wasted energy. Cutting energy alone might relax accuracy, risking misalignment during battery insertion. The only path to Pareto optimality was co-design.
Solving this problem demanded more than gradient descent. The parameter space was mixed: some variables continuous (l₁, ε), others discrete (gear ratio candidates: 34/21, 33/22, …, 30/25; screw leads: 5, 6, 8, 10, 12 mm). Moreover, evaluating each candidate required a full Simulink simulation of the lift cycle—computationally expensive. The team turned to Quantum-Behaved Particle Swarm Optimization (QPSO), an evolutionary algorithm inspired by quantum probability waves. Unlike classical PSO, where particles follow deterministic velocity updates, QPSO treats each candidate design as a “quantum particle” whose position is probabilistically sampled around a global attractor—reducing premature convergence and enhancing exploration in rugged, multimodal landscapes.
Over 200 iterations and 10,000+ simulations, the algorithm converged on a set of parameters that defied intuition. The optimal gear ratio wasn’t the originally selected 1.50 (30:20), but a tighter 1.20 (30:25), trading raw speed for torque amplification and smoother acceleration. Screw lead jumped from 10 mm to the maximum allowed 12 mm—counterintuitive, since longer leads usually reduce mechanical advantage—but in this case, it reduced the number of motor revolutions needed for the full lift, cutting I²R losses in windings more than it increased transmission friction. Link l₁ shrank slightly (112 mm vs. 110), l₂ grew significantly (219 mm vs. 180), and l₃ extended (167 mm vs. 130), reshaping the linkage’s motion profile to minimize peak inertial loads and keep the tray’s vertical velocity more uniform.
Simultaneously, the controller softened: ε dropped from 4000 to 3959, λ from 0.30 to 0.22, k rose modestly from 1.00 to 1.59. This wasn’t “weaker” control—it was smarter control. Lower λ reduced the weight on position error derivative, curbing aggressive corrections; higher k strengthened the linear damping term, smoothing the final approach. Together, they yielded faster convergence with less overshoot.
The results stunned even the researchers. In simulation, the integrated-optimization design achieved 126.48 joules per lift cycle—an 10.93% drop from the baseline 142.00 J. More impressively, steady-state angular error fell to 0.5516 radians—a 53.68% improvement over the original 1.1909 rad. To put this in perspective: at the motor shaft, that error translates to under 2 millimeters of tray height deviation—well within the ±5 mm tolerance typical for battery guide rails.
But simulations lie. Or at least, they simplify. Real-world friction isn’t perfectly modeled. Encoder latency, cable drag, and manufacturing tolerances in link joints all conspire to degrade performance. So the team built a full-scale prototype. Using the optimized parameters, they fabricated the linkage, sourced a matched motor-gear-screw assembly, and embedded their sliding-mode controller in a real-time DSP. Over 50 consecutive lift cycles with an 80-kg load, they measured energy via a Hioki power analyzer and angular error via the motor’s 20-bit encoder.
The experimental numbers held firm: 136.32 J average energy (7.78% higher than simulation, but still a 4.1% net gain over baseline hardware) and 0.5984 rad error (8.48% above simulation, yet still 49.7% better than unoptimized). The slight deviations? Attributed to unmodeled stiction in linear guides and minor mass imbalances in the tray. Crucially, both metrics remained decisively superior to any single-domain optimization—control-only or structure-only—validating their core thesis: integration wins.
Industry reaction has been measured but intrigued. One major Chinese EV battery-swap network, operating over 1,200 stations, has initiated a pilot retrofit using the Chongqing design. Early internal benchmarks suggest a 9.2% reduction in station-level electricity use during off-peak hours—translating to roughly $18,000 annual savings per site at current commercial rates. More importantly, cycle time consistency improved, allowing tighter scheduling and 5.7% higher daily swap capacity without adding hardware.
Beyond commercial impact, the study challenges a decades-old design paradigm. In robotics—and indeed most mechatronics—structure and control are siloed: mechanical engineers hand off CAD models to controls teams, who then spend months “taming” the dynamics. This handoff inevitably loses synergies. The Chongqing approach flips the script: co-simulation, co-optimization, co-ownership from sprint zero.
Of course, the work isn’t finished. The team explicitly notes in their conclusion that they’ve only optimized the lifting subsystem. The robot’s mobility—wheeled or rail-guided base—consumes even more energy during station approach and alignment. Next-phase research will unify base motion, arm articulation, and battery-handling grippers into a single energy-minimization framework, possibly using reinforcement learning for adaptive strategies across varying battery form factors and floor conditions.
Still, the foundation is laid. And it’s built on a simple but powerful idea: in electromechanical systems, the line between what the machine is and how it moves is not just blurry—it’s artificial. Efficiency lives in the interstice. By erasing that boundary, Lin Lihong and her colleagues haven’t just improved a robot. They’ve offered a new blueprint for sustainable automation—one where every joule tells a story of coordinated intelligence, from gear tooth to control gain.
As global EV sales accelerate past 20 million units annually, the silent, repetitive motions of battery-swap robots will become as defining an image of 21st-century mobility as the gas pump was for the last. And thanks to this work, those motions may just be a little gentler on the grid—and on the planet.
Lin Lihong, Cui Jiabin, Hu Zengming, Zhang Jinwen, Li Chengyuan
College of Mechanical and Vehicle Engineering, Chongqing University, Chongqing 400044, P. R. China
Journal of Chongqing University
DOI: 10.11835/j.issn.1000-582X.2022.006