HuNavSim: Human Navigation Simulator

• Human Navigation Simulator (HuNavSim) is a ROS 2-based simulation platform that models individualized, reactive human behaviors for evaluating social robot navigation.
• It integrates a modular wrapper-manager-evaluator architecture with support for multiple simulators like Gazebo, NVIDIA Isaac Sim, and Webots.
• The system employs extended Behavior Trees and stochastic Social Force Models to enable realistic human-robot interactions and comprehensive benchmarking metrics.

Searching arXiv for HuNavSim and closely related papers to ground the article. HuNavSim, the Human Navigation Simulator, is a ROS 2-based, open-source human navigation simulator designed for human-aware robot navigation: the problem of making mobile robots behave safely, legibly, and comfortably in spaces shared with people. It is positioned not merely as a crowd simulator, but as a human-behavior simulation and evaluation framework in which simulated human agents can vary individually, react to robots, and be assessed with social-navigation metrics. The original system introduced individualized, robot-reactive human behaviors and a configurable benchmarking suite, while HuNavSim 2.0 extends that foundation with stochastic local navigation, a richer Behavior Tree vocabulary for high-level human actions and social events, broader simulator support, an RViz 2 scenario-authoring workflow, and a more flexible evaluation pipeline (Pérez-Higueras et al., 2023, Escudero-Jiménez et al., 23 Jul 2025).

1. Problem setting and conceptual role

HuNavSim addresses two recurrent problems in social navigation research. First, realistic human behavior simulation is hard: many crowd simulators model pedestrians collectively, so every agent behaves similarly, which is useful for crowd flow but insufficient for local, individual reactions to robots. Second, social navigation evaluation lacks standard metrics: different tools often expose different fixed metric sets, complicating planner comparison. HuNavSim was introduced as a response to both gaps by combining a flexible ROS 2 simulation framework, individualized human behaviors that react to robots, and a large, configurable set of benchmarking metrics (Pérez-Higueras et al., 2023).

The system is therefore framed as a benchmarking tool rather than only a simulator. Its intended use is to help researchers evaluate whether a robot is not only efficient, but also socially compliant, safe, and comfortable around humans. HuNavSim 2.0 sharpens that position by emphasizing that real human studies are expensive and hard to reproduce, and by explicitly targeting the gap between crowd-flow simulators and the individual-level, socially contingent behaviors required for human-aware robot navigation. The explicit claim that HuNavSim is not just a pedestrian simulator is central to its identity: it is designed to support the development and evaluation of robot navigation systems that must operate around people in believable, socially sensitive ways (Escudero-Jiménez et al., 23 Jul 2025).

2. ROS 2 architecture and simulator coupling

The basic architecture is organized as a set of modules that interact through ROS 2 service calls. In the original formulation, HuNavSim comprises three main parts: a simulator wrapper, the hunav_manager, and the hunav_evaluator. The wrapper connects HuNavSim to a base simulator, sends the current state of the human agents to HuNavSim, receives updated poses and states, and applies them back in the simulator. The hunav_manager is the core decision module that computes the next state of each human agent from its current state and behavior. The hunav_evaluator logs simulation data and computes selected evaluation metrics after or during the run (Pérez-Higueras et al., 2023).

HuNavSim 2.0 restates this organization in wrapper-manager-evaluator form and makes the wrapper abstraction more explicit. In each simulation step, the wrapper sends the current agent state to a HuNavSim manager, which computes the next state using the behavior logic and then updates the agents back in the simulator. A separate evaluator logs the run and computes metrics at the end, saving both simulation data and metric values. The whole tool is implemented in the ROS 2 framework, and the ROS 2 foundation together with simulator-specific wrappers is presented as the basis for portability across robotics platforms. HuNavSim 2.0 is also distributed with Docker containers and a CLI to simplify setup and scenario launching (Escudero-Jiménez et al., 23 Jul 2025).

Simulator integration is a defining feature. The earlier paper describes a wrapper that can connect HuNavSim to a base simulator such as Gazebo, Morse, or Webots, and it details a dedicated Gazebo wrapper in which a world generator reads the human-agent configuration YAML, writes the Gazebo world file, spawns humans as Gazebo actors, and uses a world plugin to update agent poses during runtime. HuNavSim 2.0 broadens explicit wrapper support to Gazebo Classic, Gazebo Fortress, NVIDIA Isaac Sim, and Webots, while preserving the same architectural principle: the base simulator handles robot and environment simulation, and HuNavSim controls the human agents (Pérez-Higueras et al., 2023, Escudero-Jiménez et al., 23 Jul 2025).

A recurrent architectural boundary is that HuNavSim mainly controls global position and orientation. The earlier paper notes that Gazebo can show human animations, but realistic full-body motion is outside scope; the simulator is centered on navigation behavior, interaction logic, and evaluation rather than on full human animation fidelity (Pérez-Higueras et al., 2023).

3. Human motion models and behavior representation

At the local navigation level, HuNavSim is built primarily on the Social Force Model (SFM) and its extension for groups. SFM treats motion as being driven by goal-directed attraction, repulsion from other agents, and repulsion from obstacles. In HuNavSim, SFM provides the baseline human locomotion model and is extended to support groups and robot interaction (Pérez-Higueras et al., 2023).

The original system already supported several distinct, robot-triggered human behaviors implemented using behavior trees in BehaviorTree.CPP. These behaviors were:

• regular: standard SFM behavior; the robot is treated like another pedestrian.
• impassive: the robot is not modeled as another human in SFM; the human treats the robot as an obstacle.
• surprised: when the robot is detected, the human stops and turns attention toward the robot.
• curious: the human abandons the current goal temporarily, approaches the robot slowly, and after a default of 30 s returns to the original goal.
• scared: the human tries to stay away from the robot, adds an extra repulsive force from the robot, and decreases maximum velocity.
• threatening: the human tries to block the robot’s path, continuously computes a goal in front of the robot, and after a default of 40 s resumes original navigation (Pérez-Higueras et al., 2023).

In HuNavSim 2.0, the major extension is the use of Behavior Trees (BTs) as a richer language for high-level human behavior. The paper characterizes the advance not as a replacement for SFM, but as a new layer above it. Human local navigation remains SFM-based, while high-level BT composition allows multi-stage, event-driven, socially plausible behavior. HuNavSim 2.0 also adds controlled noise into the SFM force-factor parameters, with three configuration modes: default parameters for repeatability, custom values, or stochastic behavior by sampling parameters from normal distributions within empirically determined ranges. The stated aim is to produce subtle but realistic variability rather than random chaos (Escudero-Jiménez et al., 23 Jul 2025).

The expanded BT vocabulary includes action and condition nodes such as GoTo, LookingAtPoint, StopAndWaitTimer, isAtPosition, LookAtAgent, ConversationFormation, isSpeaking, FollowAgent, LookingAtRobot, ApproachRobot, IsLookingAtMe, and SaySomething. These nodes can be composed into BTs that control navigation, attention, waiting, social interaction, and group behavior. The paper’s warehouse example illustrates the intended expressivity: one worker checks material at several points, alternates between navigation, looking, and waiting, later detects another worker, enters a conversation group, and follows him after being told to do so; another worker checks material, notices the robot, approaches it, pauses, resumes checking, later detects the first worker, verifies mutual attention, starts a conversation, says “follow me,” and walks away with the other worker following. The example is used to show that BTs can encode sequenced tasks, conditional branching, and socially meaningful interactions in a modular way (Escudero-Jiménez et al., 23 Jul 2025).

4. Scenario authoring and supported ecosystems

HuNavSim targets canonical social-navigation situations such as Passing, Crossing, and Combined scenarios with mixed human behaviors. The original Gazebo wrapper includes example indoor settings such as cafeteria, warehouse, and house. HuNavSim 2.0 broadens the scenario repertoire further, listing examples such as cafés, warehouses, homes, hospitals, offices, factory halls, and industrial halls, and example robots such as PAL PMB2, Jetbot, Create3, Carter, and PAL TIAGo. The design principle is that users can bring their own base simulation provided a suitable wrapper is available (Pérez-Higueras et al., 2023, Escudero-Jiménez et al., 23 Jul 2025).

The principal authoring advance in HuNavSim 2.0 is the RViz 2 workflow described as a “what-you-see-is-what-you-simulate” approach. HuNavSim 2.0 provides Qt-based panels for agent configuration and metric selection. Users can load a 2-D occupancy map, place agents interactively, set their parameters and appearances, define goals by clicking on the map, assign goals to specific agents, and automatically export the scenario to YAML plus one XML behavior tree per agent. The system can also reopen existing YAML scenarios for editing, which supports iterative experimentation. A shortcut can open Groot 2 for visual inspection or refinement of the generated BTs (Escudero-Jiménez et al., 23 Jul 2025).

This scenario workflow extends the earlier configuration model rather than replacing it. In the original system, users could define the number of human agents, each agent’s behavior, goals, and other parameters either in a YAML configuration file or through an RViz-based ROS 2 GUI panel. HuNavSim 2.0 systematizes and expands that capability into a scenario-construction pipeline that couples map-based authoring, BT generation, and editable persistence. A plausible implication is that HuNavSim 2.0 reduces the friction of moving between qualitative scenario design and quantitative benchmarking, because authoring and evaluation are co-located in the same ROS 2 workflow (Pérez-Higueras et al., 2023, Escudero-Jiménez et al., 23 Jul 2025).

5. Evaluation model and metric system

Evaluation is integral to HuNavSim’s design. The original system includes 28 metrics, drawn from the authors’ prior work, SEAN 2.0, SocNavBench, Social Force Model-based measures, and other literature reviewed by Gao et al. The evaluator can compute metrics for the whole scenario, time-series metrics at each timestep, and metrics grouped by human behavior category. This allows analysis of both final outcomes and temporal interaction structure (Pérez-Higueras et al., 2023).

The metric suite spans several categories. For efficiency and motion it includes time_to_reach_goal, path_length, and cumulative_heading_changes. For proximity to people it includes avg_dist_to_closest_person and min_dist_to_people. For proxemics it includes intimate_space_intrusions, personal_space_intrusions, social+_space_intrusions, and group versions of these metrics. For completion and collisions it includes completed, min_dist_to_target, final_dist_to_target, robot_on_person_collisions, person_on_robot_collisions, and time_not_moving. Additional groups cover robot dynamic behavior, pedestrian motion, and social-force-based measures such as social_force_on_agents, social_force_on_robot, obstacle_force_on_agents, obstacle_force_on_robot, and social_work (Pérez-Higueras et al., 2023).

Several metric definitions are stated explicitly. The robot trajectory length is

$\text{path length} = \sum_i \|p_{i+1} - p_i\|$

and cumulative heading change is

$$\text{cumulative heading changes} = \sum_i |\theta_{i+1} - \theta_i|.$$

Average distance to the closest person is defined by

$\frac{1}{T}\sum_t \min_j d(r_t, p_{j,t}),$

while proxemic intrusion metrics are percentages of time inside a region around a person or group:

$$\text{intrusion \%} = \frac{\sum_t I_t}{T} \times 100.$$

The paper also states that social_work is the sum of social force on the robot, obstacle force on the robot, and social force on the agents, making it a combined measure of social disruption (Pérez-Higueras et al., 2023).

HuNavSim 2.0 extends this evaluation layer while retaining its open-ended character. Because human-aware navigation lacks consensus on benchmark scenarios and metrics, the evaluation system is described as deliberately open and extensible. The paper states that the suite grows from 28 metrics to 32 metrics after adding two new “danger and surprise” metrics from Singamaneni et al. It also adds a control mechanism through ROS services so that the user can decide when metric computation starts and stops, which is useful when only a specific phase of an experiment should be measured rather than the whole simulation (Escudero-Jiménez et al., 23 Jul 2025).

6. Empirical validation, version evolution, and limitations

The original validation study compares three planners: DWB, a standard ROS 2 navigation stack local planner optimized for goal reaching without social awareness; SCL, which uses a Social Costmap Layer and adds cost around people based on distance and orientation; and SFW, the authors’ Social Force Window planner, which predicts human motion with SFM, includes social work in trajectory scoring, and considers people’s distances, velocities, and directions. The reported findings are differentiated rather than uniformly favorable: DWB achieves the fastest and shortest paths but is the least socially aware and tends to pass closer to people; SCL increases distance from people and improves some proxemic metrics but can struggle in dense crossing situations because costs accumulate around the robot; SFW gives the best social compliance overall, performs best or near-best on social-force and social-work metrics, maintains low intimate-space intrusion, and often slows down or stops to avoid disturbing people (Pérez-Higueras et al., 2023).

The behavior-specific results are also important. In the combined scenario, regular agents behave as expected, while curious and threatening agents produce very different interaction patterns; the threatening agent is described as especially intrusive and capable of forcing the robot to nearly stop. This is presented as one of HuNavSim’s central strengths: planners can be evaluated not only under abstract crowd motion, but under distinct human-robot interaction styles (Pérez-Higueras et al., 2023).

HuNavSim 2.0 is best understood as a systematic extension of that initial platform rather than a redesign. The paper explicitly attributes six development and evaluation benefits to the new version: reducing dependence on expensive human-subject studies, increasing behavioral realism through stochastic local navigation and BT-driven social actions, supporting multiple simulators, enabling richer benchmarks through an extensible and comprehensive metric set, speeding up scenario design and iteration, and improving reproducibility and portability via ROS 2, Docker, and wrapper-based integration (Escudero-Jiménez et al., 23 Jul 2025).

Aspect	HuNavSim	HuNavSim 2.0
Simulator support	Gazebo-centered architecture; Gazebo wrapper provided	Gazebo Classic, Gazebo Fortress, NVIDIA Isaac Sim, and Webots
Behavior layer	regular, impassive, surprised, curious, scared, threatening	Extended BT action/condition vocabulary and stochastic SFM parameters
Authoring and evaluation	YAML or RViz-based ROS 2 GUI panel; 28 metrics	RViz 2 “what-you-see-is-what-you-simulate”; YAML + XML BT export; 32 metrics

The open-source status of the platform is explicit. The original paper states that both the simulator and Gazebo wrapper are publicly available at https://github.com/robotics-upo/hunav_sim and https://github.com/robotics-upo/hunav_gazebo_wrapper, respectively. The same paper also records several limitations: body animations in Gazebo are still limited; realistic full-body motion is beyond scope; current human motion is still centered on SFM; the metric set, though broad, is not exhaustive; and the reported evaluation is mainly a comparative benchmark rather than a large-scale user study with real humans (Pérez-Higueras et al., 2023).

A persistent misconception is to treat HuNavSim as interchangeable with a generic pedestrian-flow simulator. The two HuNavSim papers argue against that reading. Their emphasis on individualized behavior, robot-triggered reactions, BT composition, proxemic and force-based metrics, and planner benchmarking indicates a narrower and more technical purpose: HuNavSim is a simulation and evaluation environment for social navigation benchmarking in ROS 2, intended to support quantitative comparison of navigation methods in shared human-robot spaces (Pérez-Higueras et al., 2023, Escudero-Jiménez et al., 23 Jul 2025).