Robots can skip recalculating gravity—if their energy equations cooperate

A robot controller may be able to stop recalculating gravity as a mechanism moves and still keep an underactuated machine near its target. The trade-off is striking: the controller uses gravity’s value only at the destination while operating, but proving that this is safe still takes a detailed model of how gravity, inertia and motion interact (Harandi and Namvar, 2026).

That distinction matters in machines with fewer motors than degrees of freedom. A two-link robot arm driven at only one joint, for example, cannot directly command every movement. The joints and links tug on one another; gravity adds another force that changes with posture. Rather than controlling every coordinate independently, an engineer can shape the system’s total energy so the desired pose behaves like a stable resting point. This paper asks whether that energy-shaping strategy can work while compensating gravity only at the desired equilibrium.

The result is not a new plug-and-play recipe for every robot. It is a set of mathematical conditions under which such a recipe can be built: a certain vector field must be integrable, and an adjusted potential-energy landscape must have positive curvature near the target. The authors also show how a reformulation can replace one difficult energy-matching equation with a homogeneous one and add a design freedom that may ease the conditions. Their examples are analytical, not experimental demonstrations, so the practical promise is a route to simpler online computation—not yet evidence of faster, safer or cheaper machines.

The Science

A mechanical system stores energy in motion and position. In a robot, kinetic energy comes from moving links; potential energy includes gravity and any springs. A controller can influence the system by changing how these energies behave, much as a bowl’s shape determines whether a ball rolls toward a stable low point. This is the intuition behind energy shaping.

The paper works within interconnection and damping assignment passivity-based control, usually shortened to IDA-PBC. The method specifies a desired closed-loop mechanical system with a chosen energy function, then searches for a controller that makes the actual system follow it. Damping drains energy, while the shaped potential creates a preferred equilibrium. For underactuated systems, the search is constrained: motors can push only along actuated directions, and the remaining motions must be stabilized through the system’s natural coupling.

In conventional IDA-PBC, the control law can involve the configuration-dependent gravity term—the gradient of potential energy, which changes as the machine changes pose. The alternative studied here compensates gravity only at the desired equilibrium point. In plain terms, the controller uses the gravity force calculated for the target posture, rather than continually evaluating the current posture’s gravity force online.

The authors impose an important structural restriction: the input matrix must be a constant coordinate-selection matrix. That fits many mechanical benchmarks with direct joint torque inputs, but excludes more general systems in which the way actuators affect motion changes with configuration. The paper is therefore about a meaningful class of underactuated machines, not all robots or mechanical systems.

There is also a subtle mathematical obstacle. Once the controller stops compensating gravity throughout the trajectory, the shaped energy alone may no longer decrease in the familiar way. The authors construct a Lyapunov function—an energy-like quantity used to certify stability—by adding a correction term. That correction exists only when a vector field is integrable: roughly, its components must fit together as the gradient of a single scalar function. The corrected energy also needs a strict local minimum at the target, expressed as a positive-definite Hessian, or matrix of second derivatives.

The analysis establishes sufficient conditions for local asymptotic stability. “Local” is doing real work here: the guarantee applies near the chosen equilibrium, not necessarily from every possible initial posture. Asymptotic stability means that nearby states converge toward the target over time, subject to the paper’s damping or detectability assumptions. It does not establish a global safety guarantee or a particular region of attraction.

The authors further draw on a prior IDA-PBC reformulation that replaces the usual potential-energy matching equation with a homogeneous partial differential equation (PDE). A PDE matching condition is a constraint on a function and its derivatives across the configuration space; solving one can be a major part of controller design. The homogeneous version removes a forcing term from that equation, while an extra design term provides more flexibility. This does not remove the need to understand gravity or the system model: it shifts much of the burden to offline synthesis and verification.

What They Found

The central result is a stability theorem with two routes. In one, the potential energy separates into actuated and unactuated contributions. In the other, the designer uses the homogeneous potential-matching PDE. Both routes require the correction field to be integrable and the augmented potential to be positive near the desired equilibrium. With adequate damping—or a suitable condition ensuring that undamped motions cannot persist—the target is locally asymptotically stable (Harandi and Namvar, 2026).

A further extension adds a free design term to the control law. That term can be selected to help make the correction field integrable. It gives the designer another lever, rather than guaranteeing that a solution will always exist. For a class of two-degree-of-freedom systems, the authors turn the integrability requirement into ordinary differential equations for functions of one coordinate. Those equations offer a more manageable design test, but the paper does not claim a closed-form solution for every admissible system.

The Translational Oscillator with a Rotational Actuator, or TORA, system provides the clearest worked example. It consists of a translational oscillator and a rotational actuator, with the actuator affecting the oscillator through mechanical coupling. For a target rotational position, one construction chooses the desired inertia matrix to match the original inertia matrix. That choice makes the integrability condition straightforward, and the required stiffness parameter must satisfy \(\kappa>-k_{2}\cos(q_{2d})\) for the stated local positivity condition.

A second TORA construction targets the zero configuration and uses a different desired inertia matrix. The authors solve the integrability condition as an ordinary differential equation, obtaining \(\beta=(k_{1}\alpha/k_{2})\sec(q_{2})\). They then give a shaped potential that satisfies the homogeneous matching PDE and show that its local positivity condition holds for \(\kappa>0\). The example makes the paper’s main design point tangible: changing the desired inertia and using the homogeneous equation can make the construction easier than directly solving the conventional nonhomogeneous potential equation.

The first TORA construction also makes the limits visible. It chooses a separable potential and a particular desired inertia, rather than demonstrating a universal design. The second uses a more involved inertia choice and an explicit solution whose validity is local in configuration; the secant expression, for example, is not defined where \(\cos(q_{2})=0\). Analytic solvability in a benchmark is evidence that the conditions can be met—not evidence that they are easy to meet in an industrial robot.

The paper’s second example considers a class of two-degree-of-freedom systems with cosine-shaped potential terms, structures related to benchmark mechanisms such as the Acrobot and Pendubot. For either choice of which coordinate is actuated, the free design term is expressed through functions of the second coordinate. Enforcing integrability produces a set of ODEs. This is a constructive way to organize the search, but matching, the condition that the extra term vanish at the target, and positive curvature still need to be checked.

There are no measured performance comparisons, robot trials or numerical datasets in the supplied paper text. The reported results are symbolic conditions and analytic constructions. So charts of success rates, energy savings or computational speed would be inventions, not a digest of the evidence. The genuine numerical information available is limited to parameter inequalities in the examples; the compact displays below summarize those exact reported conditions, not experimental outcomes.

TORA stiffness condition at a general target

Label	Value
Required lower bound	0 kappa > -k2 cos(q2d)

• Required lower bound 0 kappa > -k2 cos(q2d)

The paper reports the inequality κ > −k₂ cos(q₂d); it gives no numerical parameter values.

TORA stiffness condition at zero configuration

Label	Value
Required lower bound	0 kappa > 0

• Required lower bound 0 kappa > 0

The paper reports κ > 0, not a measured performance value.

The general theorem also accommodates a nonseparable potential through a modified correction field, provided that field remains integrable and the corresponding augmented Hessian is positive definite. That extension broadens the mathematical formulation, but it does not erase the core challenge: finding a suitable desired inertia and controller terms so that all conditions hold together.

Why This Changes Things

The practical attraction is computational architecture. If a controller need not evaluate configuration-dependent gravity at every control update, it may simplify the online control calculation. That could matter in systems with constrained onboard computing, or in designs where reducing the online model evaluation is useful. But the paper does not quantify the savings, and a modern controller may find gravity calculations inexpensive. The value therefore depends on the machine, processor and competing control design.

The key nuance is that “gravity compensation at the target” does not mean “gravity is irrelevant elsewhere.” Gravity still shapes the physical motion away from the target. The proposed controller’s stability proof depends on the full model through matching and integrability conditions, and the authors explicitly say offline synthesis and verification can require the full gravity model. This is a redistribution of computation and design effort, not an escape from physical modeling.

For engineering, the most consequential contribution may be the way the paper reframes a hard design problem. Traditional energy matching can involve a nonhomogeneous PDE. The homogeneous alternative and its extra design term create room to choose a desired inertia or correction that makes integrability tractable. In the TORA example, this turns the construction into a sequence of explicit choices and conditions. For some designs, that could make controller synthesis more systematic.

Yet a theorem’s conditions can be a bottleneck. Integrability is not automatic; a vector field that looks plausible component by component may not be the gradient of any scalar energy. The constant-input-matrix assumption also narrows the class of applicable systems. And local stability says little about what happens after a large disturbance, near actuator limits, or where the model is inaccurate. The paper neither compares attraction regions with conventional IDA-PBC nor claims this method is universally superior.

That restraint is important. In robotics, a controller must work amid friction, flexible links, sensor noise, changing payloads and imperfect parameter estimates. The supplied analysis assumes a model and establishes mathematical stability under specified conditions. It does not test robustness to those real-world complications. A reduced online gravity calculation is potentially helpful, but it cannot by itself establish safer operation or lower energy use.

The broader significance is methodological: the paper expands the menu of ways engineers can shape energy in underactuated machines. Underactuation is common wherever designers want fewer actuators, lower mass, reduced cost or simpler mechanisms—but it makes control harder because some movements can only be influenced indirectly. Better synthesis tools may help engineers explore designs that would otherwise be rejected as too difficult to control. Whether this particular tool delivers that benefit remains an empirical question.

What’s Next

The next step is to implement the constructions in simulation and hardware, then compare them with conventional IDA-PBC and desired-gravity-compensation approaches. Useful measurements would include computation time per control update, tracking and settling behavior, response to disturbances, and the size of the initial-condition region from which the system converges. None of those outcomes is reported here, so they are the right questions for follow-up rather than conclusions to assume.

Researchers would also need to test how sensitive the integrability and positivity conditions are to uncertain masses, friction and gravity parameters. The analytical examples show how to solve design conditions for selected structures; they do not establish a general algorithm that finds a controller whenever one exists. Broader input mappings and systems with configuration-dependent actuator directions are another important frontier, since the current constant coordinate-selection assumption rules them out.

For practitioners, the immediate lesson is neither “gravity no longer needs modeling” nor “a lighter controller is ready.” It is more precise: in certain underactuated mechanical systems, gravity compensation can be limited online to its value at the target while a carefully augmented energy function still certifies local convergence. The proof may require substantial offline knowledge, and the design conditions may be demanding. If future experiments show that the reduced online calculation matters in real hardware, this framework could help make energy-based control more practical for machines that cannot simply power every joint independently.