At a glance
A research spotlight on Riccardo Zanella, a postdoctoral researcher at the University of Twente. His work spans energy-aware robot safety, passivity-preserving reinforcement learning, and robotic manipulation of cables and wiring harnesses.
A robot that can disentangle three tangled automotive wiring harnesses in real time is a genuinely useful machine. Riccardo Zanella is more interested in a quieter claim about that machine: that it can't, at any moment, inject more energy into its surroundings than its internal budget allows.
That idea — energy as the currency of both safety and control — runs through six publications from Zanella and his collaborators between 2023 and 2026. The work spans the rules governing physical human-robot contact. It also includes a method for teaching robots new tasks without letting a learning algorithm turn physically unstable, plus the perception and manipulation of cables, ropes and wiring harnesses. The through-line matters now because collaborative robots are leaving their safety cages, and machine-learned controllers are entering factories and homes faster than certification can keep pace.
For anyone who isn't a roboticist, the practical stakes are easy to state. The same logic that stops a learning robot from flailing dangerously while it experiments is the logic that could let a robot hand over a cable on an assembly line or work beside a person with no fence between them.
The researcher and the idea
Zanella is a postdoctoral researcher in the Robotics and Mechatronics group at the University of Twente in Enschede, the Netherlands. He earned a PhD in biomedical, electrical and system engineering at the University of Bologna in 2021, stayed there as a postdoc until early 2024, then moved to Twente. His stated interests are compact: energy-aware, performance-optimized control for safe human-robot interaction, and blending AI and optimization with physical knowledge for robots in unstructured environments.
The intellectual setting matters. The Twente group has long championed energy-aware robotics, an approach associated with co-author Stefano Stramigioli that treats physical energy flows as information about safety, about performance and about how efficiently a given motion uses energy. That framing is the spine of Zanella's program.
Where the safety numbers come from
The most recent paper sits closest to policy. In a critical review published in Results in Engineering, Zanella, Federico Califano and Stefano Stramigioli unpick the safety constraints used in physical human-robot interaction, or pHRI. They don't ask whether the numbers in the standard are right or wrong. They reconstruct where the numbers come from and which assumptions hold them up, then ask what those assumptions cost in practice.
The reference point is ISO/TS 15066, a technical specification for collaborative robots. Its power and force limiting mode sets maximum allowable forces and pressures for contact with different parts of the body. The authors walk through the derivation: once you know the maximum tolerable biomechanical load for a body region — roughly, the threshold where pain begins — you work backwards to constrain the robot's speed and force.
What makes the review distinctive is its insistence that energy is the unifying metric. As the authors note, ISO/TS 15066 in its power and force limiting mode already connects biomechanical pain thresholds to energy transfer during transient contact. Framing safety in energetic terms lets a robot monitor and modulate its energy flows in real time, using a measure that is both interpretable and directly linked to potential harm. The review, available as an open preprint, also quantifies how much performance common approximations give away.
The authors are equally candid about the foundations. The technical specification itself acknowledges that its limit values derive from a restricted set of experimental data and are subject to revision as more studies arrive. Newer work cited in the review distinguishes pain onset from maximum tolerable pain and reports statistically characterized thresholds across body regions instead of single deterministic limits.
One example shows how a single assumption ripples through the maths. The allowable impact velocity depends on a reduced-mass combination of robot and human inertia. The authors work through the extremes: a clamped body segment that cannot recoil is the conservative case, which widens the safety margin but slows the robot. Lighter links and configurations with lower reflected inertia, by contrast, allow higher safe speeds without crossing tissue force thresholds. The standard itself does not explicitly consider the clamped case.
Teaching robots without letting them blow up
Machine learning has made robot controllers dramatically more capable, but it has a certification problem. A reinforcement learning policy — a controller trained by trial and error against a reward — can, in principle, produce actions that push a physical system toward instability. Passivity-based control offers the opposite trade: strong mathematical guarantees and a reputation for sacrificing performance.
A system is passive if it doesn't generate more energy than it absorbs. Think of a bank account that can never be overdrawn. The robot can spend only energy it has been given or has stored, and when the balance reaches zero, the controller is unplugged.
The mechanism Zanella and colleagues use is the virtual energy tank, a software construct that keeps a running scalar energy budget alongside the real robot. It stores information about the physical energy flows of the system. If the tank has energy, the robot's desired action goes through. If the tank is empty, the action is cut off, which formally preserves passivity.
In Learning passive policies with virtual energy tanks in robotics, published in IET Control Theory and Applications, Zanella, Gianluca Palli, Stramigioli and Califano merge that architecture with reinforcement learning. The agent is free to discover a policy for a complex task, and the tank passivizes whatever it learns. A key quantity is the task energy: the minimum energy the tank must hold at the start to complete the task passively. If a task can be done with a finite task energy, the tank can enforce passivity while the learned policy does its job.
“The whole appeal is that the control input can be anything — it doesn't have to mimic a physical system,” Zanella said. “As long as the task energy is finite, a passive implementation is possible, and that's what makes the input a natural decision variable for optimization.”
Validation remains in simulation, which is the most obvious gap. The authors are explicit that practical application of energy tanks at its current stage remains rather limited, and they connect this to the long-running claim that passivity-based techniques lose performance. A companion conference paper, presented at the European Robotics Forum and published in Springer's proceedings, extends the idea to discrete-time tanks and stresses a practical benefit: keeping passivity during training, so even the messy exploration phase of real-world learning is guarded against hazardous, unbounded energy generation.
An independent robotics safety researcher who reviewed the work but was not involved offered a cautious read. “A tank that empties guarantees the controller stops injecting energy, but it says nothing about whether the task is actually completed well. The simulations are encouraging; the step to hardware is where performance and safety have to be balanced honestly.”
Cables, ropes and wiring harnesses
The third strand of Zanella's work is the least glamorous and possibly the most commercially immediate: getting robots to see and shape deformable linear objects, or DLOs — cables, wires, ropes and elastic tubes. These objects are everywhere in homes and factories, yet robotic systems that handle them are rare because a flexible, self-occluding, high-dimensional object is hard to perceive.
In RT-DLO, published in IEEE Transactions on Industrial Informatics, the team treats the scene as a graph. After a neural network separates cable pixels from the background, the method samples nodes along the objects' center lines and reasons about topology — proximity and orientation — to connect those nodes into separate instances. RT-DLO runs faster than 30 frames per second on a 640 by 360 image and beat both DLO-specific and general-purpose baselines on accuracy and speed, improving intersection-over-union by 2.9 percentage points with a 7 frames-per-second gain over the prior state of the art.
Where RT-DLO sees the cable, a 2024 paper in IEEE Robotics and Automation Letters models it. A neural network learns to imitate an analytical model of cable dynamics across a wide range of physical parameters, then works online to pick manipulation actions and simultaneously tune those parameters from what it observes. In experiments, the robot shaped ropes into U, A, S and I target forms on cloth and cardboard, stopping when error fell below one centimetre. The estimated bending parameters identified the black rope as roughly three times stiffer than the others — a small sign the adaptation was learning real physics rather than fitting noise. Slippage on cardboard made the parameter estimates more variable, which the authors read as evidence of more complex contact dynamics.
The newest manipulation paper, in IEEE Transactions on Automation Science and Engineering, tackles branched objects: deformable multi-linear objects such as wiring harnesses with forks and connectors. Graph neural networks learn, from a binary mask, which sampled nodes should be connected and what type each node is. Offline, the method scored above 90% on a Dice metric for predicting edges and above 90% accuracy at identifying branch and intersection points, at more than 20 frames per second. Online, a dual-arm setup used the representation to disentangle three automotive wiring harnesses.
Zanella's role in the DLO papers is the connective tissue. The manipulation work, led by Alessio Caporali and Gianluca Palli at Bologna, supplies the perception and modeling layer; the passivity work supplies the guarantee that whatever a learned controller decides to do with a cable, it cannot dump unplanned energy into the world.
The case for caution
Viewed from a critical distance, the program has a clear maturation path and equally clear boundaries. The passivity-plus-learning results are simulation-based. The ISO/TS 15066 limits they scrutinize rest on a restricted experimental dataset. The graph-based perception methods are evaluated on specific benchmark scenes and a handful of real harnesses. None of this is a flaw in the papers, which state these limits plainly; it does mean the safety claims are proofs of concept rather than deployment-ready certification.
Some experts remain cautious for a different reason: the performance question. The authors themselves call the claim that passivity-based control loses performance “highly debated,” and bolting an energy tank onto a learned policy can make an agent more conservative. Whether energy budgets constrain robots too tightly for fast, contact-rich tasks is an open empirical question that simulations won't settle alone.
What the next decade could look like
The next five to ten years will test whether energy can do for robot certification what it already does for physics: serve as a common, auditable ledger. If tank-based passivity moves from simulation onto collaborative arms, and if graph-based cable perception matures into wire-harness assembly lines, the result would be robots that are provably safe in contact and genuinely useful in the flexible, unstructured world outside the cage. Zanella's six papers sketch that path without overselling it. For a field that trades in guarantees, that restraint is itself the point.
Photo by Francesco Ungaro on Unsplash
- reinforcement-learning
- university-of-twente
- riccardo-zanella
- energy-aware-robotics
- safe-human-robot-interaction

