---
kind: concept
slug: digital-twin
title: Digital twin
summary: >-
  A software copy of a physical system or process, used to predict, test options and in some designs to control. Five uses in the sources: training a navigation policy, a six-level ship scale, edge replicas of users, a fluid-solver surrogate and an arm model for contact force. Only the arm model runs on a real robot, and only to compute torques.
tags: [simulation, reinforcement-learning, autonomous-systems, maritime, federated-learning, edge-computing, reduced-order-modeling, fluid-dynamics, force-estimation]
papers: ["2306.06766", "2411.03465", "2011.09902", "2609.17101", "2509.17053"]
related: [reward-design, proximal-policy-optimization, hierarchical-reinforcement-learning, sim-to-real-transfer, robot-simulation, domain-randomization, model-predictive-control, federated-learning, multi-agent-reinforcement-learning, dimensionality-reduction, surrogate-modeling, dynamic-mode-decomposition, imitation-learning, impedance-control]
status: draft
updated_at: 2026-11-01
---

## Overview

A digital twin is a virtual representation of an asset or process. It is fed by data and by simulators, and it is used for prediction, optimisation, monitoring, control and decisions [arxiv:2411.03465 §2]. The sources agree on that broad description but not on what the twin is for. In the radio variant and the capability scheme, the twin is used before the real system acts: a policy is trained in it, or what-if runs are made on it [arxiv:2306.06766 §III-B] [arxiv:2411.03465 §2.5]. In the edge variant, the twin is a data-backed replica of each user, and it is placed on servers that train a shared model. Its agents are trained in a simulated environment [arxiv:2011.09902 §2] [arxiv:2011.09902 alg. 1]. In the force-estimation variant, the twin runs next to the real arm and supplies a model-based torque at each control step, so it is in the loop on hardware [arxiv:2509.17053 §4.2]. Variants differ in how much they depend on live data from the real system and in whether the twin only informs the operator or also closes a control loop [arxiv:2411.03465 §2.1] [arxiv:2411.03465 §2.6].

The five sources in this wiki use the term differently:

- **As a training simulator.** A wireless digital twin is a simulated copy of an indoor environment with radio propagation. A reinforcement-learning agent interacts with it online [arxiv:2306.06766 §II-A] [arxiv:2306.06766 §III-B].
- **As a capability scale.** A digital twin of an autonomous surface vessel is sorted into six levels, from a standalone 3D model to a closed-loop autonomous twin [arxiv:2411.03465 §2].
- **As an edge-hosted replica of users.** Each user's twin holds a behaviour model, its static data and a real-time state. Twins sit on base stations, which train a shared model and decide where each twin is hosted [arxiv:2011.09902 §2] [arxiv:2011.09902 §3].
- **As a reduced-order model of a simulator.** A data model is built from the snapshots of a fluid solver and gives cheap predictions of the same flow. The card describes no feed from a physical system [arxiv:2609.17101 §2] [arxiv:2609.17101 §5].
- **As a dynamics model of a robot arm.** A MuJoCo model of the arm is updated with measured joint states, and its inverse dynamics give the torque the arm would need with no contact [arxiv:2509.17053 §4.2].

## Variants

### Radio-propagation twin for policy training

The twin is built from a map and a radio model. The mesh of receiver points is computed with ray tracing [arxiv:2306.06766 §II-A]. The agent gets radio inputs (path SNR, angle of arrival and angle of departure) along with vision and pose estimates [arxiv:2306.06766 §II-B]. The paper adds three physics terms to the PPO reward: link-state monotonicity, following the angle of arrival of the strongest path, and an SNR-gradient term where the path is fully blocked [arxiv:2306.06766 §III-B] [arxiv:2306.06766 eq. 10]. The paper says this shaped reward is not potential-based, so the optimal policy under it differs from the shortest path [arxiv:2306.06766 §III-B]. Card: [Digital Twin-Enhanced Wireless Indoor Navigation](/papers/2306.06766).

Author-reported results: on six unseen maps, the policy's success rate in 2+-NLOS is 1.0 on maps P and T, where end-to-end RL reaches 0.40 to 0.45. The heuristic baseline WAN reaches 0.85 and 0.90 on those maps [arxiv:2306.06766 tab. 4]. Normalised path length in 2+-NLOS is 2.46 to 2.82, against 4.50 to 5.14 for end-to-end RL and 3.63 to 4.01 for WAN [arxiv:2306.06766 tab. 3]. The ablation that replaces the SNR term with a relative-distance cost raises 2+-NLOS path length to 4.62 [arxiv:2306.06766 tab. 5]. The tables do not name the maps used in the ablation.

### Capability-level twin for autonomous vessels

The levels run from 0 to 5 [arxiv:2411.03465 §2]:

1. Level 0, standalone: a 3D geometric model and simulations, possibly with no data link [arxiv:2411.03465 §2.1].
2. Level 1, descriptive: a one-way data stream from the asset, with fusion and estimation of quantities that are hard to measure [arxiv:2411.03465 §2.2].
3. Level 2, diagnostic: condition monitoring and cybersecurity monitoring [arxiv:2411.03465 §2.3].
4. Level 3, predictive: forecasts of future states for maintenance and process optimisation [arxiv:2411.03465 §2.4].
5. Level 4, prescriptive: "what-if" simulation to choose an operating strategy [arxiv:2411.03465 §2.5].
6. Level 5, autonomous: a two-way link in which the twin receives data and controls the asset in a closed loop [arxiv:2411.03465 §2.6].

The authors adapt the scheme from earlier work and say the boundaries are judgement calls [arxiv:2411.03465 §2]. For each level they list enabling technologies: sensors and filters at level 1 [arxiv:2411.03465 §4.2], condition monitoring at level 2 [arxiv:2411.03465 §4.3], and path-following and collision-avoidance controllers at level 5, including MPC, reinforcement learning and hybrids [arxiv:2411.03465 §4.6.1]. Card: [Digital Twin for Autonomous Surface Vessels](/papers/2411.03465).

The review runs no experiments and reports no new measurements [arxiv:2411.03465 §3]. Its numbers, such as a 4 to 6 % fuel saving from online estimation, are quoted from other works and not tested [arxiv:2411.03465 §4.1.2].

### Edge-hosted user twins for federated learning

Each user has a twin, and each twin is linked to at most one base station. The link depends on the twin's data size, the station's compute capacity and the transmission rate [arxiv:2011.09902 §2.1] [arxiv:2011.09902 eq. 1]. Each base station averages the models of its own twins, and a macro base station averages the station models [arxiv:2011.09902 §2.2] [arxiv:2011.09902 eq. 4] [arxiv:2011.09902 eq. 5]. The stations are the nodes of a permissioned blockchain that uses delegated proof of stake, and each station's model is recorded as a transaction for the others to verify [arxiv:2011.09902 §2.3].

The paper casts the choice of association, batch sizes and bandwidth as a problem of minimising total time to a target global accuracy. It calls this combinatorial [arxiv:2011.09902 §3] [arxiv:2011.09902 eq. 18]. The solver is a multi-agent DDPG with one agent per base station. All agents receive the same reward, the negative of the total time, because the slowest station sets that time [arxiv:2011.09902 §4.1] [arxiv:2011.09902 eq. 19]. Card: [Low-latency Federated Learning and Blockchain for Edge Association in Digital Twin empowered 6G Networks](/papers/2011.09902).

Author-reported results: the text gives no numbers. The experiment uses CIFAR-10 with 100 users, five base stations and one macro station, an IID split and a small CNN [arxiv:2011.09902 §5]. The figures show lower total time cost than random and average association [arxiv:2011.09902 fig. 5], loss comparable to full-data federated learning [arxiv:2011.09902 fig. 6], and a discount factor of 0.9 converging fastest among the values tested [arxiv:2011.09902 fig. 7]. The experiment has no repeated runs, no non-IID data and no measured blockchain overhead [arxiv:2011.09902 §5].

### Reduced-order data model of a simulator

The paper's "digital twin data model" is a surrogate for a numerical solver, not a copy of a physical asset. Snapshots come from a finite-difference solver for the 1D viscous Burgers equation at Reynolds numbers 10², 10³ and 10⁴, with 301 snapshots on 101 grid points [arxiv:2609.17101 §2]. A randomized SVD inside dynamic mode decomposition (DMD) gives a low-rank model. A rank search chooses the number of modes by minimising a time-averaged error [arxiv:2609.17101 §4] [arxiv:2609.17101 alg. 1] [arxiv:2609.17101 eq. 29]. A cascade-forward network with 10 hidden units, called an NLARX estimator, then predicts the time series of each mode's amplitude [arxiv:2609.17101 §5] [arxiv:2609.17101 fig. 3] [arxiv:2609.17101 eq. 32]. The pair forms the data model [arxiv:2609.17101 eq. 31]. Card: [randomized-DMD data model for fluid flows](/papers/2609.17101).

Author-reported, all in-sample: the DMD error is between 5.4e-8 and 4.4e-7 at the three Reynolds numbers, with correlation 1.0 for all three [arxiv:2609.17101 tab. 1]. The group-level data-model error is between 1.3e-4 and 8.1e-4 [arxiv:2609.17101 tab. 2] [arxiv:2609.17101 tab. 3] [arxiv:2609.17101 tab. 4]. The online stage takes 8 to 17 s and the offline stage at most about 2 s, on hardware the paper does not state. No full-order solver time is given [arxiv:2609.17101 §6] [arxiv:2609.17101 fig. 13]. The paper does not state which rank-search method produced Table 1 [arxiv:2609.17101 §6].

### Joint-torque twin for contact force estimation

A MuJoCo model of the arm runs in real time next to the robot. The measured joint positions, velocities and torques are copied into it [arxiv:2509.17053 §4.2]. The twin's inverse-dynamics torque, the torque the joints need to hold the pose with no contact, is subtracted from the measured torque. A Jacobian pseudo-inverse maps the difference to a six-axis wrench at the end effector [arxiv:2509.17053 §4.2] [arxiv:2509.17053 eq. 6]. The estimate replaces a wrist force sensor as an input to the imitation policy [arxiv:2509.17053 §1] [arxiv:2509.17053 §4.1]. Card: [FILIC: Dual-Loop Force-Guided Imitation Learning with Impedance Torque Control for Contact-Rich Manipulation Tasks](/papers/2509.17053).

Author-reported results: in simulated peg insertion, the policy with estimated force reaches 90.0% success, against 68.0% with end-effector position only and 80.0% with joint torque as input [arxiv:2509.17053 tab. 2]. On four real tasks, estimated force reaches 76.6 to 83.3%, against 26.6 to 46.7% for position only and 60.0 to 70.0% for joint torque [arxiv:2509.17053 tab. 3]. The paper gives no numbers for the accuracy of the force estimate itself. The static load test is shown only as a figure [arxiv:2509.17053 §5.1] [arxiv:2509.17053 fig. 4]. The paper says the twin accounts for sensor noise, friction and mechanical compliance [arxiv:2509.17053 §4.2]. The released estimator does not model these; see the card's "Paper vs code" section before relying on the twin for them.

## When to use which

| Situation | Use | Why |
|---|---|---|
| You must train a policy before touching the real system, and the target environment is mapped | Radio-propagation twin with a shaped reward | The agent learns by interacting with the twin, and the paper's tests are on unseen simulated maps [arxiv:2306.06766 §V] |
| You do not have a propagation model or a map of the environment | Not covered by any source | The first variant needs both to build the twin [arxiv:2306.06766 §II-A] |
| You accept a policy that is suboptimal for the original cost in exchange for better transfer | Physics-shaped reward | The paper says the shaping changes the optimum and argues it helps transfer [arxiv:2306.06766 §III-B] |
| You need to decide which sensing, monitoring and control to build for an asset | Capability levels as a checklist | The scheme names enablers per level, with no architecture or validation [arxiv:2411.03465 §4] [arxiv:2411.03465 §5] |
| You need forecasts or "what-if" runs to choose operations, without closing the loop | Levels 3 and 4 | These levels use prediction and simulation but not control [arxiv:2411.03465 §2.4] [arxiv:2411.03465 §2.5] |
| You want the twin to control the real asset | Level 5, two-way link | The review describes the closed loop [arxiv:2411.03465 §2.6], but gives no experiment that tests it |
| Many user devices must be assigned to edge servers under compute and bandwidth limits, and you can train the assignment in simulation | Edge-hosted user twins with multi-agent DDPG | The paper trains one agent per station and reports lower time cost than two simple association rules [arxiv:2011.09902 §4] [arxiv:2011.09902 fig. 5] |
| You need evidence that the assignment works on real hardware or non-IID data | Not covered by any source | The results are figures only, with one IID dataset and no variance [arxiv:2011.09902 §5] |
| You need a tamper-evident record of shared models among base stations | Blockchain ledger with delegated proof of stake | The design is given, but stake, rewards and producer counts are not stated [arxiv:2011.09902 §2.3] |
| You need a cheap surrogate of a solver you can sample densely, and in-sample error is enough for now | Reduced-order DMD twin with a neural amplitude model | The paper reports low errors on three Burgers cases, but all in-sample [arxiv:2609.17101 tab. 1] [arxiv:2609.17101 tab. 2] |
| You need the surrogate to hold for new initial conditions, Reynolds numbers or time windows | Not covered by any source | No held-out window, initial condition or Reynolds number is tested [arxiv:2609.17101 §6] |
| You need contact force for a policy, cannot fit a wrist sensor, and have an accurate arm model | Joint-torque twin with pseudo-inverse wrench | The paper reports gains over position-only inputs on a robot it does not name [arxiv:2509.17053 tab. 3] [arxiv:2509.17053 §5.4] |
| You need the force estimate to be accurate under friction and compliance | Not covered by any source | The paper claims the twin handles them but gives no accuracy figure for the estimate [arxiv:2509.17053 §4.2] [arxiv:2509.17053 fig. 4] |

The strongest evidence for a training twin is simulated. The two baselines in [arxiv:2306.06766 tab. 4] are close to the proposed method on some maps, so a team with a working heuristic should compare against it. The edge paper's baselines are two simple association rules, not other federated or blockchain schemes [arxiv:2011.09902 §5]. The force-estimation paper compares input types on the same policy, not against other force-sensing methods [arxiv:2509.17053 §5.4].

## Implementation sketch

Training a policy in a radio-propagation twin with a physics-shaped reward. The shape of the loop follows the paper; the code is pseudocode. The edge variant's training loop is in [arxiv:2011.09902 alg. 1] and is not repeated here.

```pseudocode
# Train a navigation policy against a radio-propagation twin  [arxiv:2306.06766 §II-A, §IV]
# Input:  twin (map, radio mesh), reward weights, policy pi
# Output: trained policy parameters

def train_in_twin(twin, pi, weights, n_episodes):
    # Sample a map and a transmitter position  [arxiv:2306.06766 §V]
    for episode in range(n_episodes):
        pose, radio = twin.reset()
        done = False
        while not done:
            # Global policy picks a heading and a link-state estimate  [arxiv:2306.06766 §IV-A]
            heading, link_state = pi.act(radio)
            # The twin returns the next radio state from its mesh  [arxiv:2306.06766 §II-A]
            pose, radio, done = twin.step(heading, link_state)
            # Physics terms are added to the distance cost  [arxiv:2306.06766 eq. 10]
            cost = twin.distance_cost(pose)
            cost += weights.ls * twin.monotonicity_penalty()
            cost += weights.aoa * twin.aoa_gap(heading)
            cost += weights.snr * twin.snr_gradient_gap(heading)
            # Update the policy from the shaped reward  [arxiv:2306.06766 §IV-B]
            pi.update(-cost)
    return pi
```

The paper's own weight search covers [0.5, 2] [arxiv:2306.06766 §IV-B]. Use the same range as a starting point and fix it before comparing against a baseline.

## Contradictions

No numerical conflicts: the edge paper reports only figures, and the other sources share no numbers. The sources do use the term with different definitions. The review defines a twin as enabled by data and simulators [arxiv:2411.03465 §2]. The radio paper uses it as a simulator for training [arxiv:2306.06766 §II-A]. The edge paper uses it for a replica of each user's state on an edge server [arxiv:2011.09902 §2]. The reduced-order paper uses it for a model of a solver's output, with no feed from a physical system [arxiv:2609.17101 §2]. That case does not fit the review's data-fed definition, so its use of the name should not be read as support for the review's levels. The force-estimation paper is the one source where the twin is updated from a real machine during operation [arxiv:2509.17053 §4.2]. Its use is closer to the review's data-fed sense, but the review's levels do not describe it.

Points worth noting when reading a single source:

- The review defines a twin as data-enabled, yet allows Level 0 with no data link [arxiv:2411.03465 §2] [arxiv:2411.03465 §2.1].
- The reduced-order paper's errors are in-sample only, and its correlation metric is 1.0 in every case, so it does not separate the cases [arxiv:2609.17101 §6] [arxiv:2609.17101 tab. 1].
- The training twin's reported numbers come from a reward that the card for that paper says the released code does not implement as written. Check the card's "Paper vs code" section before attributing the numbers to the released code [arxiv:2306.06766 §IV-B].
- The force paper's twin is said to model noise, friction and compliance, and its accuracy is not reported as a number [arxiv:2509.17053 §4.2].

## Papers

- [Digital Twin-Enhanced Wireless Indoor Navigation](/papers/2306.06766): a radio-propagation twin for training a PPO navigation policy, with physics terms in the reward.
- [Digital Twin for Autonomous Surface Vessels](/papers/2411.03465): a six-level capability scheme for vessel digital twins, with a survey of enabling technologies per level.
- [Low-latency Federated Learning and Blockchain for Edge Association in Digital Twin empowered 6G Networks](/papers/2011.09902): edge-hosted user twins, a blockchain for shared models, and a multi-agent DDPG for association and bandwidth.
- [Randomized-DMD data model for fluid flows](/papers/2609.17101): a DMD-based reduced model of a Burgers solver, with a neural estimator of its mode amplitudes; in-sample results only.
- [FILIC: Dual-Loop Force-Guided Imitation Learning with Impedance Torque Control for Contact-Rich Manipulation Tasks](/papers/2509.17053): a MuJoCo twin that turns joint torques into an end-effector wrench for a force-conditioned imitation policy.