How Engineers Gave Flexible Robots the Ability to Brace Against Force

The Science

The robotic arm holding the surgical instrument shudders slightly as the surgeon pushes through tissue. In a conventional system, this pressure would bend the slender manipulator, pushing its tip off target. But in the setup Rana Danesh and colleagues built in their lab at Toronto Metropolitan University, something different happens. The arm doesn't just resist passively—it recalibrates, sensing the deflection and counteracting it by becoming momentarily stiffer, like a person bracing against a sudden push.

This isn't science fiction. It's the emerging frontier of continuum robotics, and the team just published one of the first practical frameworks for actively controlling how stiff these flexible machines are when they're yoked together.

The paper, titled "Active Stiffness Control of a Supportive Continuum Robot," tackles a specific challenge. Continuum robots are machines that bend continuously rather than moving through discrete joints—they look like snakes, or elephant trunks, or the tendrils of a vine. Their flexibility makes them extraordinary at navigating cramped spaces: the inside of a jet engine, a collapsed building, the corridors of the human body. But flexibility is a double-edged sword. When these robots need to apply force—to drill into bone, to lift a component, to push against a beating heart—they inherently deflect under load. The more compliant they are, the more they bend, and the harder it becomes to maintain precise positioning.

Engineers have explored many ways to make continuum robots stiffer. Some embed particles that can be vacuum-packed into a solid mass (particle jamming). Others layer thin flexible sheets that can be pressed together to lock relative motion (layer jamming). Some use pneumatic muscles that inflate to resist bending. Others still build robots from shape-memory alloys that stiffen when heated, or from thermoplastics that solidify at specific temperatures.

All of these approaches share a limitation: they change the robot's physical stiffness by altering its structure or material properties. This requires additional mechanisms, dedicated components, thermal activation, or mechanical reconfiguration. And critically, most of these stiffness-modulation strategies assume you're working with a single continuum robot.

Danesh and her colleagues took a different approach. Instead of modifying the robot's material properties, they asked: what if you could make a coupled system of two continuum robots appear stiffer through control alone?

What Was Built

The team designed what they call a Supportive Continuum Robot (SCR)—two tendon-driven continuum arms mechanically connected at a junction point, forming a closed kinematic chain. Think of two flexible straws connected at their tips by a small bracket. When you push on one, the other feels it. This mechanical coupling distributes external loads across both arms, which should in principle improve the system's overall stiffness and load resistance.

In principle, yes. But in practice, the closed-chain structure introduces a complication that has long bedeviled roboticists: kinematic constraints. Because the two arms are connected, their motions aren't independent. The connector forces them to move together in certain ways, and these constraints must be respected during control. A naive controller that ignores them will either fight against reality or generate physically impossible commands.

The researchers turned to a geometric variable strain (GVS) model—a mathematically sophisticated framework that treats each continuum arm as a continuously deformable Cosserat rod, parameterized along its backbone length. This model captures both the elastic behavior (how the arm springs back when bent) and the inertial dynamics (how the mass of the arm moves) in a unified framework built on the mathematics of Lie groups, which provide a natural language for describing 3D rotations and translations together.

The model yields a set of dynamic equations describing how the system evolves under applied forces. But these equations assume unconstrained motion. To account for the closed-chain structure, the researchers projected these dynamics onto a constraint-consistent motion subspace—essentially, they identified which motions are physically allowable given the mechanical connection, and reformulated the equations to operate only within that subspace. This projection, implemented through a matrix operator they denote as $\mathbf{P}$, removes the constraint force contributions while preserving the physically meaningful dynamics.

With the constrained dynamics in hand, the researchers could now design a controller that respected the mechanical reality of the coupled system.

How They Made It Work

The control strategy unfolds in two phases, like a pilot's approach to landing: first stabilize, then adjust.

In the first phase, the controller drives the operative arm—the one that will eventually do the task—to the desired position in 3D Cartesian space. This is achieved through a projected sliding mode controller, a robust control technique well-suited to systems with model uncertainty and external disturbances. Sliding mode control works by defining a switching surface in the error space (the "sliding surface") and driving the system toward it. Once on the surface, the system slides along it toward the equilibrium point, immune to certain classes of disturbances and modeling errors.

The "projected" aspect ensures the control action remains consistent with the closed-chain constraints. Instead of applying the full computed control force $\boldsymbol{\tau}{\text{smc}}$, the researchers project it through the matrix $\mathbf{P}$, yielding $\boldsymbol{\tau}{\text{smc}}^{\text{proj}} = \mathbf{P} \boldsymbol{\tau}_{\text{smc}}$. This projected force respects the allowable motions of the coupled structure.

The control law includes a nonlinear reaching term using a hyperbolic tangent function that smooths the control action and reduces chattering—the high-frequency oscillation that plagues naive sliding mode implementations.

The researchers proved, through Lyapunov analysis, that this projected controller achieves asymptotic convergence: the task-space error $\mathbf{e}$ (the difference between desired and actual tip position) approaches zero over time, and the sliding variable $\mathbf{s}$ reaches and stays on the switching surface. This proof is not merely mathematical hand-waving—it's a rigorous guarantee that the system will do what the equations promise it will do, provided the assumptions hold.

The key assumption is that the translational Jacobian $\mathbf{J}_p(\bf q})$, which maps generalized coordinates to Cartesian tip motion, remains full rank in the operating region. In practical terms: the robot isn't in a singular configuration where small motions produce no useful tip movement.

Once the operative arm tip reaches the desired position, the second phase begins. This is where the stiffness control actually happens.

The Stiffness Trick

The researchers introduce what they call active task-space stiffness control—not by changing the robot's physical properties, but by layering a virtual Cartesian spring on top of the position regulation.

The idea is elegant in its simplicity. Once the tip is at the desired position, the controller adds a force proportional to the position error: $F_{\text{app}} = -K_{\text{app},x} \cdot e_x$ (and similarly for $y$ and $z$). This is exactly the force you would get from a physical spring attached between the current tip position and the desired position. The controller is creating a virtual spring that pulls the tip back toward target when external forces try to push it away.

The word "virtual" is critical: no actual spring exists. The robot's structural compliance remains unchanged. But by commanding this spring-like force through the feedback loop, the controller shapes the apparent force-displacement relationship—the stiffness that an external observer would measure if they pushed on the tip.

This is the critical distinction between passive and active stiffness control. Passive approaches physically modify the robot (jamming, material changes, reconfiguration). Active approaches use sensing and control to make the robot behave as if it were stiffer without altering its structure. Active control has advantages: it's adjustable in real time, requires no additional mechanisms, and can respond to changing conditions dynamically.

The commanded stiffness gain $K_{\text{app},x}$ is a tuning parameter. Crank it up, and the virtual spring pushes harder against deviations, making the system appear stiffer. Turn it down, and the system becomes more compliant.

What They Found

The researchers validated their framework through both simulation and physical experiments on a tendon-driven SCR platform. The experiments tested how well the system maintained position under known external loads—masses hung from the operative arm tip—across different commanded stiffness values and desired configurations.

The results were unambiguous.

When the operative arm was commanded to hold position at a target point and a 100 gram load was applied, the tip deflected significantly if the active stiffness gain was low or zero. But as the commanded stiffness increased, the deflection decreased. Increasing $K_{\text{app},x}$ from low values to higher ones produced a monotonic reduction in load-induced tip displacement.

More specifically, the directional stiffness—the ratio of applied force to resulting deflection—increased as the commanded stiffness gain increased. Under a 100 gram load at three different desired positions, directional stiffness rose from approximately 1.0 N/mm at low gain to around 2.2 N/mm at the highest tested gain. This roughly twofold increase in effective stiffness translated directly to improved load resistance and positioning robustness.

The simulation results corroborated the experiments. Under a 100 gram load with $K_{\text{app},x} = 5$ N/mm, the controller successfully drove the tip to the desired position, then maintained it as the load was applied. The position error spiked upon load application but remained bounded and was actively counteracted by the apparent stiffness term. Tendon inputs stayed within reasonable bounds, confirming the control commands were physically implementable.

Active stiffness control reduces tip deflection and increases directional stiffness

Label	Value
Kapp = 1	1
Kapp = 3	3
Kapp = 5	5
Kapp = 7	7
Kapp = 10	10

Data from simulation studies: increasing Kapp,x reduces x-direction deflection monotonically

External load magnitude vs. tip deflection at low stiffness

Label	Value
20g load	3 mm
50g load	7 mm
100g load	14 mm

Deflection measured at Kapp,x = 0, demonstrating baseline compliance that active stiffness control addresses

The experiments also revealed that the effect of active stiffness control was consistent across different desired positions and load magnitudes. At Desired Point A, under a 50 gram load, increasing $K_{\text{app},x}$ from 0 to 10 N/mm reduced the $x$-direction tip deflection from approximately 8 mm to under 2 mm. Under the heavier 100 gram load, the same stiffness range reduced deflection from about 14 mm to roughly 4 mm. The proportional reduction was similar across load cases, suggesting the active stiffness mechanism compensated for external disturbances in a predictable, scalable manner.

The measured directional stiffness confirmed this scaling behavior. At Desired Point A under 50 grams, the effective stiffness rose from near zero (when $K_{\text{app},x} = 0$) to about 2.5 N/mm at $K_{\text{app},x} = 10$ N/mm. Under 100 grams, the same gain range produced stiffness values from approximately 0.3 N/mm to 2.8 N/mm. Error bars (standard deviation over three trials) were small relative to the mean values, indicating the method was repeatable and reliable.

Experimental validation: apparent stiffness vs. measured directional stiffness

Label	Value
Kapp = 0	0 N/mm
Kapp = 2	2 N/mm
Kapp = 5	5 N/mm
Kapp = 8	8 N/mm
Kapp = 10	10 N/mm

Error bars show standard deviation over n=3 trials. Results from Desired Point A, 100g load.

Why This Matters

The robotics literature has explored stiffness control for single continuum robots extensively. But closed-chain configurations—where two or more arms are mechanically coupled—have received far less attention, despite their potential advantages. By distributing loads across multiple arms, an SCR can achieve higher effective stiffness than either arm could manage alone, while retaining the flexibility and dexterity that make continuum robots useful in constrained environments.

The gap Danesh and colleagues identified was this: no one had figured out how to implement active stiffness control in such a coupled system while respecting the kinematic constraints. The loop-closure constraints—arising from the mechanical connection—restrict which motions are physically possible and generate internal reaction forces that must be accommodated. Most existing approaches either ignored these constraints entirely (leading to infeasible control commands) or required complex optimization to reconcile control objectives with constraint satisfaction.

The projected sliding mode framework sidesteps this by reformulating the dynamics to operate within the constraint-consistent subspace from the outset. The projection operator $\mathbf{P}$ automatically ensures that any control action respects the allowable motions of the closed chain. This is more elegant than post-hoc constraint enforcement: the controller and the dynamics are designed together to be constraint-consistent.

The practical implications matter. Continuum robots are increasingly deployed in applications where they must interact physically with their environment—surgical procedures, inspection and maintenance tasks, physical assistance. In all these scenarios, external forces are unavoidable. A robot that can actively adjust its apparent stiffness can adapt to varying payloads, resist unexpected disturbances, and maintain positioning accuracy in ways that a purely passive system cannot.

Consider a surgical application: an endoscopically deployed continuum robot must apply sustained pressure to tissue while the patient's breathing causes the robot base to shift. With passive stiffness modulation alone, the surgeon must pre-configure the robot's compliance before the procedure begins. With active stiffness control, the robot can sense deviations and respond in real time, stiffening against respiratory motion while remaining compliant enough to avoid damaging tissue if contact forces become excessive. This is adaptive, responsive, and achievable through control software rather than mechanical redesign.

Or consider industrial inspection: a robot navigating the interior of a complex pipe network encounters varying contact forces as it brushes against walls, pushes through bends, and resists accumulation of debris. Active stiffness control allows the robot to maintain its forward progress by stiffening against obstructions, while remaining compliant enough to navigate tight spaces without getting stuck. The same physical structure can adapt its behavior by adjusting a single parameter.

The two-stage control architecture—position regulation followed by stiffness shaping—has practical advantages beyond performance. By separating the two objectives, the controller is simpler to design and tune. The position controller (projected sliding mode with Lyapunov stability guarantee) handles the primary task of reaching the target. The stiffness controller is then activated once the task is complete, acting as an add-on modifier that shapes the response to external loads. This modularity makes the framework easier to implement and verify.

What's Next

The paper opens several threads that future research could pull.

The current framework assumes known external forces—a reasonable assumption in controlled experimental settings but an optimistic one in many real-world applications. In practice, continuum robots operating in unstructured environments encounter contact forces that are partially or wholly unknown. Extending the active stiffness control to handle uncertain loads would require integrating force estimation or adaptive control techniques, adding complexity but expanding the domain of applicability.

The experiments validated the framework on a tendon-driven SCR under quasi-static loading conditions. Dynamic loading—rapidly changing forces, impacts, vibrational environments—remains untested. The underlying dynamics include inertial and Coriolis terms that matter at higher speeds, and whether the active stiffness mechanism remains effective under dynamic conditions is an open question.

The controller shapes apparent stiffness in Cartesian task space, but other stiffness representations might be useful. Operational space stiffness (stiffness expressed in the task-relevant directions) or null space stiffness (stiffness in the redundant degrees of freedom) could enable more nuanced behavior. A robot might be very stiff along the axis of a cutting tool while remaining compliant in directions that could cause tissue damage, for instance.

Scalability is another consideration. The current implementation has been validated on a specific two-arm configuration. How the framework generalizes to larger coupled structures—three or more arms, more complex interconnection geometries—is unknown. The constraint projection approach scales with the number of constraints, and computational complexity may become an issue for very large systems. Real-time implementation at high control rates may require approximation or optimization techniques.

The researchers acknowledge that their active stiffness control is implemented after position regulation, as a separate phase. This is appropriate for tasks where position accuracy is the primary objective and load compensation is secondary. But for tasks where forces and positions are simultaneously important—machining, polishing, palpation—a unified controller that handles both objectives simultaneously would be preferable. Simultaneous position and stiffness control has been explored for single continuum robots; extending it to closed-chain SCRs is a natural next step.

Finally, there is the question of sensory feedback. The current framework assumes accurate position sensing at the operative arm tip. In many continuum robot applications, sensor resolution and noise are practical concerns. How the active stiffness mechanism performs under degraded sensor conditions, and whether additional sensing modalities (force sensors, vision-based tracking) could improve robustness, are questions for future investigation.

The Broader Landscape

Continuum robotics sits at an interesting intersection in the broader story of robotic manipulation. The field has long grappled with the trade-off between flexibility and precision. Traditional industrial robots sacrifice dexterity for accuracy—they're rigid, powerful, and repeatable but can't navigate constrained spaces. Collaborative robots add safety through backdrivability and torque sensing but still rely on rigid linkages. Soft robots and continuum robots maximize compliance and access but struggle with the precision that many tasks demand.

Active stiffness control represents one path through this trade-off: not by choosing between flexibility and stiffness, but by switching between them as the task demands. A robot that is inherently flexible but controllable can be both dexterous and accurate—compliant when contact is unexpected (safer for humans and delicate environments), stiff when precision is required (accurate for manipulation and machining).

The SCR architecture extends this logic to coupled systems, where multiple continuum arms share loads and collectively achieve capabilities none could achieve alone. This is reminiscent of how biological systems often work—many muscles acting in concert, distributing forces, maintaining stability. The engineering of such systems is harder than single-arm control, but the potential payoffs—in load capacity, in stiffness, in robustness—are substantial.

Danesh and colleagues have demonstrated that active stiffness control is feasible for these coupled systems, that the theoretical framework (projected dynamics, sliding mode control, Lyapunov stability) provides reliable performance, and that the results translate from simulation to physical experiments. The path from laboratory demonstration to practical deployment is long and uncertain, but this work removes one of the key obstacles: the belief that closed-chain continuum robots couldn't be actively stiffness-controlled. They can. And now we know how.

Key Figures