Imagine a marble rolling inside a smooth bowl. No matter where you release the marble, gravity and friction will eventually conspire to bring it to rest at the very bottom. The bottom of the bowl is a natural, asymptotically stable equilibrium point because it represents the state of minimum potential energy.
Now imagine you want to control a complex physical system—such as a robotic arm, an aerial drone, or an electric power converter. Traditional control theory often treats this as a mathematical problem of error cancellation: you calculate the difference between where the system is and where it should be, and then apply massive external forces to fight the system's natural physics.
Energy-shaping control takes the opposite approach. Instead of fighting the physics, it rewrites them. By viewing a system through the lens of its total energy (its Hamiltonian or Lagrangian), an energy-shaping controller mathematically modifies the system's internal energy landscape so that the desired target state becomes the new "bottom of the bowl."
Reshaping the Energy Landscape
To change a physical system's behavior, we can inject virtual potential and kinetic energy through its control inputs. In a mechanical system, total energy is the sum of kinetic energy (motion) and potential energy (position). By modifying these functions, we can alter how the system naturally moves.
In Energy Shaping Control of a CyberOctopus Soft Arm, Heng-Sheng Chang et al. demonstrate this concept by controlling a highly flexible, continuous soft robotic arm. Instead of calculating complex inverse kinematics for every segment of the flexible body, they modify the potential energy function within the controller so that the target configuration becomes the global minimum. The arm then naturally relaxes into the goal state without the chaotic oscillations typically associated with high-degree-of-freedom continuous structures.
This methodology applies far beyond robotics. For instance, Wei He et al. apply Interconnection and Damping Assignment Passivity-Based Control (IDA-PBC) to stabilize DC-DC buck-boost converters in An Adaptive Passivity-Based Controller of a Buck-Boost Converter With a Constant Power Load. By defining a target closed-loop energy function that accounts for the converter's inherent non-linearities, the system settles at the desired voltage without relying on fragile linear approximations.
Even when moving from continuous physics to digital implementations, the structural integrity of this approach holds. Anthony M. Bloch et al. show in Controlled Lagrangians and Potential Shaping for Stabilization of Discrete Mechanical Systems that deriving potential and kinetic shaping laws directly from discrete variational principles preserves the underlying physical structure, ensuring robust digital stabilization for physical hardware.
Ensuring Stability via Passivity and Damping
Reshaping the energy landscape only ensures that the target state is an equilibrium; it does not guarantee the system will stop moving once it gets there. Without a way to drain excess energy, the system will simply oscillate around the minimum forever. To achieve true stability, the system must be passive and include damping injection.
A system is passive if it cannot produce more energy than it absorbs. By enforcing passivity, controllers guarantee that a system will never become unstable or generate unexpected behaviors. To force the system to settle at the energy minimum, designers inject virtual damping to dissipate energy. For example, Fadil Santosa demonstrates in Control of the Wave Equation by Time-Dependent Coefficient that monitoring the time rate of change of potential energy provides an ideal feedback signal for parameter switching, allowing a system to maximize energy dissipation and damp out structural vibrations.
[Reshaped Potential Energy (Goal = Minimum)] + [Damping Injection (Energy Dissipation)]
│
▼
[Asymptotic Asymptotic Stability]
Managing this energy flow is critical when dealing with real-world constraints like time delays, obstacles, or environmental interactions:
- Teleoperation and Delays: In remote robotic control, communication lag can introduce artificial energy into the control loop, causing dangerous oscillations. To counteract this, Mattia Risiglione et al. implement virtual energy observers in Passivity-based control for haptic teleoperation of a legged manipulator in presence of time-delays to monitor and neutralize spurious energy. Similarly, Michael Panzirsch et al. introduce "deflection-domain" passivity control in Enhancing the Force Transparency of the Energy-Reflection-Based Time-Domain Passivity Approach, shifting energy monitoring to the physical stretch of a virtual coupling spring to eliminate jarring force drops while maintaining strict passivity.
- Physical Interaction: When drones or robots interact with physical surfaces, active control terms can inadvertently generate destabilizing energy. Min Jun Kim et al. utilize time-domain passivity control in Passive Compliance Control of Aerial Manipulators to inject virtual damping the moment energy generation is detected. For complex environments, Carlo Tiseo et al. utilize a piece-wise energy manifold to define a fractal attractor topology in Bio-mimetic Adaptive Force/Position Control Using Fractal Impedance, ensuring intrinsic stability during contact even without a perfect mathematical model of the environment.
- Obstacle Avoidance and Assistance: Passivity can be dynamically modulated to enforce safety boundaries or provide assistance. Lukas Huber et al. develop a passive obstacle-aware controller in Passive Obstacle-Aware Control to Follow Desired Velocities that modulates damping based on obstacle proximity, guaranteeing collision avoidance through energy management. Conversely, Auke Ijspeert et al. utilize negative damping in Adaptive Negative Damping Control for User-Dependent Multi-Terrain Walking Assistance with a Hip Exoskeleton to inject energy in phase with a user's movement, minimizing resistance and maximizing power efficiency.
- Hybrid Robustness: To handle severe external disturbances, energy-shaping can be combined with other robust paradigms. Han Wang combines energy-based passive control with sliding mode techniques in Passive Sliding Mode Controlled UPFC and its Treatment Strategy of Unbalanced Grid Conditions, injecting damping through Euler-Lagrange modeling while utilizing a sliding surface to compensate for grid unbalance and modeling inaccuracies. Similarly, Andrea M Tonello et al. apply a global change of coordinates alongside IDA-PBC in Energy Based Control of a Swash Mass Helicopter Through Decoupling Change of Coordinates to decouple complex dynamics and ensure a unique unmanned aerial vehicle naturally converges to its target trajectory.
Data-Driven Energy Landscapes: Learning Port-Hamiltonian Dynamics
The traditional bottleneck of energy-shaping control is that it requires an exact mathematical model of the system's physics (such as its mass matrix, inertia, and potential energy). For complex, non-linear systems, deriving these equations by hand is incredibly difficult.
The modern breakthrough in this field solves this problem by combining energy-shaping control with deep learning via Neural Ordinary Differential Equations (Neural ODEs). Instead of training arbitrary black-box neural networks, researchers structure these networks to explicitly learn the underlying Hamiltonian or Port-Hamiltonian components of the system from raw data.
Raw Trajectory Data ──► [Port-Hamiltonian Neural ODE] ──► Learned Energy Functions ──► Passivity-Based Control (IDA-PBC)
In Symplectic ODE-Net: Learning Hamiltonian Dynamics with Control, Yaofeng Desmond Zhong et al. introduce an actuated energy-based modeling framework that bakes the laws of Hamiltonian mechanics directly into a neural network's architecture. By learning how external control forces interact with internal energy, the model extracts interpretable physical components—like mass and potential energy—directly from data.
Because these learned models strictly respect physical invariants like energy conservation, they generalize remarkably well to unseen scenarios. Nikolay Atanasov et al. extend this to rigid-body systems in Hamiltonian-based Neural ODE Networks on the SE(3) Manifold For Dynamics Learning and Control, forcing the network to respect the geometry of the $SE(3)$ manifold. The learned energy functions are then used directly in IDA-PBC laws for stable trajectory tracking without needing manually specified physical parameters.
Why does this combination work so robustly? As Thai Duong et al. explain in Port-Hamiltonian Neural ODE Networks on Lie Groups For Robot Dynamics Learning and Control, passivity-based control laws remain structurally valid even if the learned energy functions are imperfect or scaled versions of reality. The underlying port-Hamiltonian structure guarantees stability despite parameter uncertainties.
This structural guarantee allows for powerful optimization and safety synthesis:
- Optimization: Stefano Massaroli et al. demonstrate in Optimal Energy Shaping via Neural Approximators that neural networks can parameterize and optimize energy-shaping functions and damping gains simultaneously, outperforming hand-tuned controllers while guaranteeing physical safety.
- Autonomous Safety: Zhichao Li et al. show in Safe Autonomous Navigation for Systems with Learned SE(3) Hamiltonian Dynamics that by dynamically shaping a learned energy landscape, a robot can safely navigate unknown environments. The energy landscape naturally draws the robot toward its goal while ensuring it slows down near obstacles and accelerates in open spaces, maintaining strict passivity and stability throughout its flight.
Key Foundational Work
Controlled Lagrangians and Potential Shaping for Stabilization of Discrete Mechanical Systems (2006) — contributed 1 concept elements to this direction.
Control of the Wave Equation by Time-Dependent Coefficient (2002) — contributed 1 concept elements to this direction.