💻 computer science

Oncomorphic computing: a bioinspired paradigm for decentralized adaptation across software and hardware

This paper introduces "oncomorphic computing," a bioinspired paradigm that repurposes cancer survival principles—such as decentralized behavior, phenotypic plasticity, and resource recruitment—to create resilient machine intelligence capable of autonomously adapting and restructuring across software and hardware in noisy, damaged, or resource-scarce environments.

Original authors: Rosalia Moreddu, Philip Greulich, Christian Baumgartner, Francesca Borghi, Michael Levin

Published 2026-10-29

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Rosalia Moreddu, Philip Greulich, Christian Baumgartner, Francesca Borghi, Michael Levin

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a machine that does not break when parts of it fail, but instead learns to live with the damage, shifting its goals and reorganizing its own functions to keep working. This is the central challenge facing the next generation of autonomous systems, from spacecraft traveling through deep space to medical implants inside the human body. Current artificial intelligence is incredibly powerful, but it relies on a foundation of stability: reliable power, constant communication, and hardware that does not degrade. When these conditions vanish, as they often do in extreme environments or during long-term biological monitoring, standard systems tend to collapse. They are built to be perfect, and when they are not, they stop. To solve this, researchers are looking to an unlikely source for a blueprint: the biology of cancer. While cancer is a disease of uncontrolled growth, it also represents a masterclass in decentralized survival. Tumors persist without a central brain, adapt to starvation, and reorganize their internal structure when parts are destroyed. By studying how cancer cells survive and thrive in hostile conditions, a team of scientists has proposed a new way to design machines that can do the same.

The researchers, a group of experts from universities in the United Kingdom, Austria, Italy, and the United States, have introduced a concept they call "oncomorphic computing." This is not a single device or a specific type of software, but a new philosophy for building intelligent systems. The core idea is to stop treating machines as rigid structures that must be repaired or replaced when they break. Instead, they propose designing systems where every small part, whether it is a software program, a sensor, or a robot joint, acts like an independent agent. These agents constantly monitor their own health and their local environment. If a part of the system is damaged or running out of power, the agents do not wait for a central computer to issue a fix. They negotiate with their neighbors, sharing resources and shifting responsibilities. If one agent can no longer perform its job, it might pass its tasks to a healthier neighbor, or the whole system might simply decide to drop a less important function to save energy for the most critical ones.

The paper argues that this approach is necessary because the future of autonomy lies in unpredictable places. A rover on Mars cannot wait for instructions from Earth if a solar flare disrupts its communication. A robot swarm in a disaster zone cannot rely on a central server if the network is fragmented. A digital model of a patient's tumor must be able to adapt as the disease changes, even if the data coming in is incomplete or noisy. In these scenarios, the goal is not to be perfectly accurate at every moment, but to remain useful for as long as possible. The authors suggest that by borrowing principles from cancer biology—such as the ability of cells to change their identity, recruit resources from neighbors, and migrate functions to safer areas—engineers can build machines that are resilient by design.

To make this idea concrete, the researchers have created a mathematical framework that describes how these systems should behave. They imagine a network of agents, each with its own memory, power, and processing ability. In a healthy state, these agents work together to maximize performance. But when stress hits, the system changes. Damaged agents might be isolated to prevent them from spreading errors, while healthy agents absorb their tasks. The system does not try to fix the broken part immediately; instead, it reorganizes around the damage, preserving the most important mission goals. This is a shift from the traditional way of thinking, where a machine is expected to execute a specific program exactly as written. In this new view, the machine is a living network that negotiates its own survival, constantly deciding which functions are worth keeping and which can be let go.

The paper explores how this concept could be applied across different fields. In software, it suggests breaking down large artificial intelligence models into smaller, independent modules that can adapt to local failures. If a part of a robot's brain fails, the rest of the system can reassign tasks to keep the robot moving. In hardware, the idea points toward new types of computer chips that can rewire themselves when parts are damaged, similar to how a tumor reorganizes its blood supply. The researchers also look at the biological side of the equation, suggesting that this same framework could help scientists better understand cancer itself. By treating a tumor as a decentralized computing system, they can simulate how it adapts to drugs and stress, potentially leading to better treatments.

One of the most striking applications discussed is in the realm of digital medicine. The authors propose using this approach to create "digital twins" of human tumors. These would be computer models that mimic the behavior of a specific patient's cancer, composed of thousands of virtual agents representing individual cells. As new data comes in from the patient, the model updates itself, learning how the tumor is likely to evolve. Because the model is built on the same principles of adaptation as the real tumor, it could predict how the disease will respond to different treatments, even in the face of uncertainty. This could help doctors choose therapies that are more likely to work, not just based on the average patient, but based on the unique, chaotic behavior of that specific individual's disease.

The researchers are careful to note that this is a proposal for a new way of thinking, not a finished product. They acknowledge that building such systems is difficult. It requires balancing the energy cost of constant adaptation with the need for stability. If a system spends too much energy reorganizing itself, it might run out of power before it can do its job. There are also safety concerns. A machine that can change its own structure and goals must be carefully controlled to ensure it does not make dangerous decisions. The paper suggests that these systems need built-in safety layers that can override the adaptive process if things go too far, ensuring that the machine never loses sight of its core mission.

Despite these challenges, the potential rewards are significant. The paper suggests that by embracing the messy, decentralized nature of biological survival, we can build machines that are far more robust than anything we have today. These systems would not just survive damage; they would use it as a signal to reorganize and continue their work. Whether it is a robot exploring a cave, a sensor monitoring a patient's heart, or a computer modeling a complex disease, the goal is the same: to create intelligence that can endure. The authors believe that this shift in perspective, from building perfect machines to building resilient ones, is essential for the future of technology. They are not just proposing a new algorithm or a new chip; they are proposing a new way to think about what it means for a machine to be alive, adaptable, and useful in a world that is constantly changing.

The journey from this idea to reality will take time. The researchers outline a path forward, starting with simulations and small-scale tests, moving toward more complex hardware, and eventually reaching full-scale robotic and biological applications. They see a future where the line between software and hardware blurs, where the physical structure of a machine is as adaptable as its code. In this future, a machine that loses a part does not stop; it simply becomes a different kind of machine, one that is still capable of doing what matters most. This is the promise of oncomorphic computing: a vision of technology that does not fear failure, but learns from it, just as life has done for billions of years.

Technical Summary: Oncomorphic Computing

1. Problem Statement

Contemporary artificial intelligence and autonomous systems are predominantly designed for stable, resource-rich environments characterized by reliable hardware, consistent communication, and ample power. However, critical applications—such as deep-space exploration, implantable medical devices, distributed sensor networks, and robotic swarms—operate under conditions of noise, damage, scarcity, and unpredictability. In these environments, standard fault-tolerant approaches (redundancy, rollback, error correction) often fail because they assume failures are known, isolable, or recoverable via predefined mechanisms. When degradation is sudden, resources are competitive, and communication is delayed, existing systems struggle to maintain reliability and longevity. The core challenge is designing machine intelligence that can preserve useful function and mission viability when the computational substrate itself is failing, rather than merely correcting errors within a stable framework.

2. Methodology and Framework

The paper proposes Oncomorphic Computing, a bioinspired paradigm that abstracts survival principles from cancer biology (specifically decentralized behavior, phenotypic plasticity, heterogeneous specialization, and resource recruitment) and repurposes them for software and hardware design.

2.1 High-Level Formalization

The paradigm is formalized as a constrained decentralized control problem on a time-varying graph $G(t) = (V, E(t))$ , where vertices represent local agents (software processes, hardware tiles, sensors, or cells) and edges represent dynamic pathways.

Agent State: Each agent $i$ $i$ maintains a state vector $z_i(t)$ $z_{i} (t)$ comprising:
- Task State ( $x_i$ ): Internal model weights, cached data, or biological variables.
- Health ( $h_i \in [0,1]$ ): A continuous measure of the unit's ability to contribute useful tasks, evolving under damage and stress.
- Confidence ( $\tau_i \in [0,1]$ ): A measure of output reliability, distinct from physical health.
- Resources ( $r_i$ ): Local allocable memory ( $m$ ), power ( $p$ ), and bandwidth ( $b$ ).
- Mode ( $\sigma_i$ ): A discrete state (active, relay, dormant, migrating) representing phenotypic plasticity.
Local Control: Agents operate on partial observability, using local observations $o_i(t)$ to select actions via a policy $\pi_i$ .
Adaptation Dynamics:
- Health Update: Health evolves based on accumulated damage, induced load, and compensatory support from neighbors ( $S_i$ ).
- Resource Recruitment: Agents can transfer resources ( $T_{ij}$ ) to neighbors, modeling the "inheritance" of function seen in biological systems.
- Task Viability: System success is measured by the retained utility of critical functions ( $V(t)$ ) rather than global task accuracy. The system prioritizes a "protected mission backbone," shedding non-essential tasks to preserve critical ones under stress.
Safety Mechanisms: The framework includes quarantine rules to isolate units with low integrity and external safe-mode conditions to prevent uncontrolled self-modification.

2.2 Comparison with Adjacent Paradigms

The paper distinguishes oncomorphic computing from:

Classical Digital Computing: Shifts focus from exact execution on stable hardware to preserving mission utility under substrate degradation.
Distributed/Edge AI: Moves beyond protocol-based coordination to health-aware node valuation and functional inheritance during progressive degradation.
Neuromorphic Computing: While neuromorphic systems mimic neural dynamics for efficiency, oncomorphic systems specifically address the management of physical substrate damage and resource scarcity without central control.
Swarm Intelligence: Unlike homogeneous swarms, oncomorphic agents possess heterogeneous internal states, compete for resources, and undergo role differentiation based on survival rather than just representational accuracy.

3. Key Contributions and Application Frontiers

The paper outlines how this paradigm addresses specific bottlenecks and applies to several domains:

3.1 Machine Learning Bottlenecks

Centralized Overhead: Proposes breaking monolithic stacks into interacting agents with local health estimates, allowing for resource recruitment and task simplification when communication degrades.
Catastrophic Forgetting: Treats heterogeneity as a resource. Stable modules preserve high-value capabilities while plastic modules absorb new information, preventing the overwriting of essential functions.
Scalability: Replaces global consensus with event-triggered communication and health-aware routing, enabling systems to shift into "triage" modes under severe stress.

3.2 Adaptive Hardware and Co-Design

Short-term: Utilizes reconfigurable digital hardware (e.g., FPGAs) where partial reconfiguration and fault-aware management allow for graded remapping and task migration.
Long-term: Explores memristive and in-memory computing substrates. These devices offer local storage and computation, enabling adaptation at finer spatial scales, though they require managing non-idealities like variability and non-linear programming.

3.3 Computational Biology and Digital Medicine

Digital Twins: Proposes constructing patient-specific tumor twins as populations of interacting cell agents. These twins continuously assimilate data (imaging, molecular) to update local parameters, predicting disease trajectories and resistance mechanisms.
Computational Oncology: Models tumors as distributed adaptive systems where agents integrate bioelectrical states (membrane potential, calcium dynamics) with metabolic and mechanical cues to simulate transient adaptive states (e.g., drug tolerance) that bulk models miss.
Closed-Loop Bioelectronics: Enables automated, state-triggered perturbation in in vitro systems. Sensing regions act as local agents that intervene (via stimulation or drug delivery) only when confidence is sufficient, creating a closed loop for adaptive biophysics.

3.4 Robotic Systems

Deep-Space Exploration: Reorganizes control around local viability, allowing rovers to reduce terrain ambition or shift roles (e.g., from explorer to relay) based on energy and link quality.
Humanoid Robotics: Implements graded self-preservation where control authority migrates to the most competent body parts (e.g., shifting load from a thermally stressed knee to the hands) as local competence changes.
Soft Robotics: Provides an arbitration layer for systems where damage alters mechanics and sensing simultaneously, allowing the system to decide which regions are reliable enough to govern behavior.
Surgical Robotics: Creates a "live trust field" where every sensing route maintains its own state, allocating authority based on current confidence rather than nominal capability.
Wearable Sensors: Treats reliability as distributed, negotiating trust and sampling density across a network of sensors to maintain physiological monitoring despite interface drift or physical damage.

4. Results and Claims

The paper does not present new experimental data but rather a conceptual framework and a roadmap for implementation.

Theoretical Contribution: It provides a mathematical formalization for a system where "viability" is the primary objective, distinct from "accuracy."
Roadmap: The authors propose a staged development path:
1. Near-term: Software demonstrations (neural agents, resource-aware continual learning) and health-aware distributed inference.
2. Intermediate: Linking policies to adaptive substrates (fault-aware FPGAs, hybrid digital-in-memory accelerators) and refining biological formalisms (Bayesian digital twins).
3. Long-term: Embodied deployment in robotics, surgical systems, and wearables capable of autonomous structural and functional adaptation.
Metrics: The paper defines specific metrics for evaluation, including Retained Mission Utility (RMU), recovery time ( $T_{rec}$ ), adaptation energy overhead ( $\eta_{adapt}$ ), and unsafe state counts.

5. Significance and Conclusion

The paper claims that oncomorphic computing represents a fundamental shift in designing resilient machine intelligence. By abstracting decentralized survival principles from cancer biology, it offers a framework for systems that can continue to function under uncertainty, scarcity, and partial failure. Its significance lies in bridging the gap between software dynamics and adaptive hardware design, moving beyond the assumption of a stable substrate. The authors argue that this approach is essential for future applications in digital health, autonomous systems, and embodied robotics, where the ability to preserve critical functions through continuous negotiation and restructuring is more valuable than maintaining perfect accuracy in a failing environment. The realization of this paradigm requires addressing challenges in computational overhead, hardware non-idealities, and the ethical governance of self-restructuring systems.