SUMO: Multi-Domain Simulator & Ontology

Updated 12 July 2026

SUMO is a polysemous acronym that denotes both the open-source Simulation of Urban MObility for traffic and the Suggested Upper Merged Ontology for knowledge representation.
The traffic simulator models road networks with directed graphs, calibrated flows, and integrates digital twins for precise traffic analysis.
Recent research extends SUMO into computer vision, offline reinforcement learning, and LLM optimization, enhancing performance and efficiency.

SUMO is a polysemous acronym in contemporary research. In transportation and autonomous-systems literature, it most commonly denotes Simulation of Urban MObility, an open-source traffic simulator used for microscopic and mesoscopic modeling, network generation, control, digital twins, and benchmarking. In knowledge representation and commonsense reasoning, it denotes the Suggested Upper Merged Ontology, an upper-level ontology written in SUO-KIF and widely studied through first-order-logic translations. More recent work also reuses “SUMO” for unrelated methods in computer vision, offline reinforcement learning, and memory-efficient large-language-model training (Guastella et al., 2023, Álvez et al., 2018, Tian et al., 29 Jun 2026, Qiao et al., 2024, Refael et al., 30 May 2025).

1. Principal referents of the acronym

Field	Expansion	Representative paper
Transportation simulation	Simulation of Urban MObility	(Guastella et al., 2023)
Knowledge representation	Suggested Upper Merged Ontology	(Álvez et al., 2018)
Computer vision	Segment and Track Any Motion with Nonlinear State Space Models	(Tian et al., 29 Jun 2026)
Offline reinforcement learning	Search-Based Uncertainty estimation for Model-Based Offline RL	(Qiao et al., 2024)
LLM optimization	Subspace-Aware Moment-Orthogonalization	(Refael et al., 30 May 2025)

The transportation and ontology senses are historically the most established in the supplied literature. The former appears as a simulator with a large tooling ecosystem, while the latter appears as an upper-level ontology designed to support semantic interoperability, theorem proving, and extensions into domain ontologies. The later machine-learning usages are independent acronym expansions rather than extensions of either earlier system (Liang, 20 Dec 2025, Allen, 2020).

2. Simulation of Urban MObility as a traffic-simulation platform

Simulation of Urban MObility represents a road network as a directed graph of nodes $\mathcal{N}$ , edges $\mathcal{E}$ , and connections $\mathcal{C}$ . Agents such as vehicles, cyclists, and optionally pedestrians move on this network under three parameterized behavior models: a car-following model, a lane-changing model, and an intersection model. In the Waymo benchmarking study, the default car-following model is Krauss, the lane-changing model is SL2015, and the intersection model enforces right-of-way and traffic-signal compliance (Liang, 20 Dec 2025).

A canonical workflow begins with network generation or import. Abstract networks can be created with netgenerate, while real networks are commonly imported from OpenStreetMap through netconvert. Demand is then defined by vehicle types, trips, or flows, using tools such as od2trips and randomTrips.py, and route assignment is carried out with duarouter or, for iterative user-equilibrium routing, duaIterate.py. A minimal .sumocfg bundles the network, route, and additional files, after which the simulation is run with sumo or sumo-gui. Calibration proceeds by importing traffic counts or GPS traces through laneAreaDetector or TraCI, comparing simulated and real counts, and scaling OD-matrix cells or insertion rates until simulated flows match observed flows (Guastella et al., 2023).

The simulator exposes multiple output formats for post-processing. tripinfo.xml contains departure, arrival, duration, route-length, waiting-time, and speed for each vehicle; fcd-output logs full trajectories; summary reports aggregate metrics by edge or lane; edgedata provides per-edge time series; and laneAreaDetector outputs counts, speed, occupancy, and time-loss summaries. The tutorial explicitly frames traffic analysis around the relationships $Q = k \cdot v$ , $TT_i = t_{\text{arrival},i} - t_{\text{depart},i}$ , and $\mathrm{Delay}_i = TT_i - T_{ff}$ , with $Q$ as flow, $k$ as density, and $v$ as speed (Guastella et al., 2023).

3. Tooling, interfaces, and control-oriented workflows

SUMO has accumulated a substantial ecosystem for scenario generation, calibration, and external control. One line of work uses aggregate count data to drive intersection-level simulations. In the Toronto case study, turning movement count data from the City of Toronto Open Data Portal are preprocessed from CSV or JSON, mapped to SUMO <flow> definitions, combined with an intersection-only network extracted from OSM, and validated against observations using RMSE, MAPE, and Theil’s $U$ . The stated purpose is intersection traffic generation “without creating full vehicle routes through the network,” thereby keeping network complexity to a minimum (Maheshwari et al., 14 Aug 2025).

Another line of work wraps SUMO’s command-line tools behind higher-level orchestration layers. SUMO-MCP exposes nine sub-modules, including network, route, traffic_signal, detector, visualization, and xml, through an MCP client-server architecture with dynamic import. The platform supports natural-language-triggered workflows such as OSM download, OD conversion, batch signal-plan simulation, comparative reporting, and congestion detection. In a Chaoyang evening-peak experiment with 5,000 vehicles, the reported average travel times were 845.34 s for Fixed-Time, 773.97 s for Actuated, 703.13 s for Webster, and 838.35 s for GreenWave; the platform also reports average MCP execution of 74 s and 5 calls versus 106 s and 12.6 calls for direct CLI usage (Ye et al., 4 Jun 2025).

SUMO is also commonly integrated with reinforcement learning and multi-agent control. In the Jurong EV charging framework, SUMO runs in step-by-step mode under TraCI, a Python controller reroutes a target EV to one of five charging stations, and the state vector is $\mathcal{E}$ 0, where $\mathcal{E}$ 1 are shortest-path distances, $\mathcal{E}$ 2 are en-route traffic counts, and $\mathcal{E}$ 3 are queue lengths. The action space is the charging-station index, and the terminal reward is $\mathcal{E}$ 4. Across experiments with 200, 300, and 400 simultaneous EVs, Dueling DDQN outperformed Random, Greedy, and DQN, with the heaviest scenario reported as approximately 843 s average travel time versus approximately 900 s for DQN and approximately 960 s for Greedy (Song et al., 2022).

For traffic-light control, the JADE–TraSMAPI–SUMO tool-chain places JADE on top of TraSMAPI and SUMO’s TraCI server. Traffic-light agents observe a state tuple $\mathcal{E}$ 5, select phase-duration adjustments through Q-learning, and use a reward defined as $\mathcal{E}$ 6 a weighted average of neighbors’ rewards. The reported result is that semi-fixed control outperforms a naïve fixed plan, and that Q-learning over phase durations and day period yields slight further reductions in average travel time together with improved homogeneity (Azevedo et al., 2016).

In autonomous-vehicle testing, SUMO has been coupled both to OpenAI Gym and to CARLA. SUMO-Gym packages scenario selection, TraCI lifecycle management, and observation/action handling behind a standard Gym API while populating background vehicles with sampled IDM parameters calibrated from naturalistic driving data. The CARLA–SUMO–Gym framework then runs SUMO and CARLA in lock-step at $\mathcal{E}$ 7 s, maps SUMO vehicle states into CARLA’s world frame, and renders camera, LiDAR, and radar streams for modular AV-pipeline evaluation (Kusari et al., 2021, Li et al., 2021).

A recurring technical issue in the SUMO literature concerns the fidelity of mesoscopic simulation. The mesoscopic model used in SUMO at the time of Ni et al.’s study, denoted SUMO-MESO, derives from Eissfeldt’s 2004 segment-based queue-dynamics model. Ni et al. argue that it only partially embodies first-order LWR theory because it mixes cell-transmission and link-transmission ideas, delays queue spillback, and underestimates congestion. They propose a discrete-time Link Transmission Model that tracks cumulative entering and exiting flows, link density, sending flow, receiving flow, and explicit backward-traveling spaces. In a signalized one-way corridor, SUMO-MESO under-predicted peak link densities by up to 20% and dissipated congestion about 200 s earlier, whereas the proposed LTM nearly overlapped with SUMO-MICRO density curves with RMSE below 3 veh/km. In the motorway lane-drop case, LTM stayed close to the theoretical fundamental-diagram envelope with maximum deviation about 5%, ran within 5% CPU time of SUMO-MESO, and remained an order of magnitude faster than microscopic SUMO (Ni et al., 8 Jun 2026).

At the microscopic level, realism has been extended to specific road users. The SimRa-based cyclist study replaces scalar bicycle parameters with vTypeDistribution sampling over fitted acceleration, deceleration, and maximum-speed distributions, and adds the parameter indirectLeftTurnProb for intersection behavior. The four introduced cyclist models are bicyclesAll, bicyclesSlow, bicyclesMed, and bicyclesFast, with indirect-left-turn probabilities of 0.61, 0.97, 0.87, and 0.50 respectively. Across three German cities and 300 bicycle runs per street, the new models shifted acceleration, deceleration, and speed distributions closer to observed SimRa data, while the added sampling overhead was reported as negligible, below 1% runtime (Karakaya et al., 2023).

SUMO is also used to construct calibrated digital twins and to benchmark model-based simulation against data-driven alternatives. On Dublin’s M50 motorway, a SUMO-based digital twin fused inductive loops, toll records, GPS probes, and motorway-camera outputs into complete-information and partial-information replicas. The model achieved 93.1% accuracy in average speed estimation and 97.1% in average trip length estimation; ICEV $\mathcal{E}$ 8 was overestimated by 0.8–2.4%, and EV power consumption was underestimated by 1.0–5.4% (Wang et al., 14 Jul 2025). In a separate benchmark against Waymo Open Motion Dataset scenarios, SUMO reached a realism meta metric of 0.6532 on WOSAC with fewer than 100 tunable parameters, and in 60 s rollouts recorded a collision rate of 0.0047 and an offroad rate of 0.0073, indicating stronger long-horizon stability than representative data-driven simulators (Liang, 20 Dec 2025).

Although most open SUMO scenarios have been road-traffic centric, the GROSS pipeline extends SUMO to nation-scale rail simulation. GROSS combines OSM railway infrastructure with GTFS schedules, uses topology-aware stop mapping through a hierarchical station model, performs station-level routing and targeted repair, and produces Germany-wide scenarios containing 35,925 trips. Across multiple German regions it reduces average teleportations per vehicle by factors ranging from 1.7 to 76.8, and in the North Rhine-Westphalia example it raises the share of stops with delay below 1 min from 85% to 90% while reducing median delay from 2.5 min to 2.0 min (Penell et al., 2 Jun 2026).

5. Suggested Upper Merged Ontology in knowledge representation

In knowledge representation, SUMO denotes the Suggested Upper Merged Ontology, described as one of the oldest and largest publicly available upper-level ontologies. It was originally developed under the auspices of the IEEE Standard Upper Ontology Working Group around 2001 to provide a high-coverage, language-neutral scaffold of commonsense categories and relations, including events, objects, attributes, processes, and roles. SUMO is written in SUO-KIF and, like virtually all large ontologies, is designed under the Open World Assumption (Álvez et al., 2018).

For first-order automated theorem proving, SUMO is commonly translated into pure FOL. Adimen-SUMO, cited as the most successful FOL rendition of SUMO’s two upper layers, contains roughly 8,300 formulas, of which 5,255 are atomic, and defines 2,169 classes. Its structural backbone is built around subclass, instance, and disjoint. The subclass relation is axiomatized as a partial order, including transitivity,

$\mathcal{E}$ 9

while instance(a,C) ties particulars to classes and disjoint(C,D) forbids common instances or subclasses (Álvez et al., 2018).

A major issue identified in commonsense-reasoning work is missing structural knowledge under the Open World Assumption. Over the $\mathcal{C}$ 0 million ordered pairs of classes, the ontology yields 18,374 provable subclass pairs, 3,304,246 provably disjoint pairs, and 62,069 pairs sharing an instance or subclass, but leaves 1,381,941 pairs without an asserted subclass link and 1,338,246 pairs without a disjointness verdict. The study characterizes this as nearly 30% missing structural knowledge. Applying a Careful Closed World Assumption to subclass and disjoint raises the number of solved commonsense competency questions from 7,285 of 14,324 under OWA, or 50.86%, to 9,781 of 14,324, or 68.28%, under CWA-subclass plus CWA-disjoint with assumed disjointness; counting any question solved by either CWA version increases the total to 10,970 of 14,324, or 76.58% (Álvez et al., 2018).

The WordNet–SUMO mapping is central to these experiments. WordNet synsets are linked to SUMO terms through equivalence $\mathcal{C}$ 1, subsumption $\mathcal{C}$ 2, and instance-level $\mathcal{C}$ 3 relations. A later analysis of 16,972 competency questions, based on a 1% random sample of 169 cases, reported 111 correct mappings and 58 incorrect mappings, with severe adjective-related errors in antonymy patterns because many WordNet adjectives had been linked to SUMO Process classes rather than SUMO Attribute classes. The authors recommend systematic re-evaluation of the mapping, use of incompatible competency questions as alignment tests, augmentation of SUMO’s process knowledge, and iterative benchmark-driven ontology enrichment (Álvez et al., 2019).

SUMO has also been used outside heavyweight ATP workflows for semantic simulation. Allen’s study implements a “simplified Python SUMO” that imports ontology fragments, represents classes and relations as Python objects, and treats SUMO rules as methods that assert or retract relations across state changes. The work illustrates this with gasoline-engine simulations and proposes a Generative Lexicon-style structured definition $\mathcal{C}$ 4, thereby repositioning SUMO as a model-level specification framework as well as an inference substrate (Allen, 2020).

6. Recent unrelated acronymic uses in AI and machine learning

Recent papers reuse the acronym “SUMO” for methods that are unrelated to either traffic simulation or ontology engineering. In computer vision, “SUMO” expands to Segment and Track Any Motion with Nonlinear State Space Models. The method is a zero-shot, training-free framework for visual object tracking and moving-object segmentation. Its state vector is eight-dimensional, its observation is four-dimensional, and it combines a nonlinear state-space model with a Selective Unscented Filter and a memory-selection mechanism. Reported results include LaSOT-ext AUC 61.2% versus 59.0 for the best prior baseline, GOT-10k AO 83.5%, SegTrackv2 $\mathcal{C}$ 5, FBMS-59 $\mathcal{C}$ 6, and DAVIS16-MOVING $\mathcal{C}$ 7 (Tian et al., 29 Jun 2026).

In offline reinforcement learning, “SUMO” expands to Search-Based Uncertainty estimation for Model-Based Offline RL. The method defines uncertainty through the cross-entropy between learned model dynamics and dataset dynamics and approximates $\mathcal{C}$ 8 using a particle-based $\mathcal{C}$ 9-nearest-neighbor estimator over concatenated $Q = k \cdot v$ 0 vectors, implemented efficiently with FAISS. When integrated into MOPO and AMOReL on D4RL MuJoCo tasks, the reported average normalized scores rise from 59.7 to 69.3 for MOPO and from 72.2 to 80.3 for AMOReL; runtime for MOPO+SUMO is reported as 6 h 41 m versus 7 h 23 m for vanilla MOPO (Qiao et al., 2024).

In large-language-model optimization, “SUMO” expands to Subspace-Aware Moment-Orthogonalization. The optimizer performs truncated SVD-based subspace extraction, projects the first moment into the low-rank basis, orthogonalizes the resulting $Q = k \cdot v$ 1 matrix exactly through SVD, and maps the update back to the original parameter space. The authors argue that this removes the approximation error associated with Newton–Schulz orthogonalization in ill-conditioned regimes. Reported empirical results include up to 20% memory reduction relative to state-of-the-art methods, approximately 1.6× faster convergence on QNLI than GaLore, and improved GLUE, C4 pre-training, and GSM8K results (Refael et al., 30 May 2025).

These later usages demonstrate that “SUMO” has become an acronymic label reused across fields rather than a domain-unique term. This suggests that, in technical writing, expansion on first mention is not merely stylistic but necessary for disambiguation across transportation, ontology engineering, vision, reinforcement learning, and optimization.