Independent crowd-aware navigation involves robots autonomously reaching goals in human-populated environments while minimizing disruption, as seen with service robots in malls or airports integrating into pedestrian flows (Yao et al., 2019). This includes systems designed for joining moving groups (Truong and Ngo, 2017) or avoiding stationary crowds (Tsoi et al., 2022). Independent navigation is the most widely studied and versatile form of social navigation.

Assistive navigation tasks involve robots directly supporting humans, such as follower robots in airports (Gupta et al., 2016), shopping assistants (Chen Y. et al., 2017), interactive guides (Burgard et al., 1999; Thrun et al., 2000), and systems aiding visually impaired individuals (Chuang et al., 2018), or accompanying people and groups (Ferrer et al., 2017; Repiso et al., 2020). Some tasks include proactively offering guidance (Kato et al., 2015). These tasks require detecting, following, and interpreting human cues for safe and seamless assistance.

Collaborative navigation features robots and humans working together on shared tasks, either physically or through shared control. In industry, cobots assist on assembly lines (Matheson et al., 2019), while human mobility robots use shared-control systems, model-based (Gonon et al., 2021) or learning-based (Zhang et al., 2023) to integrate human input and dynamically adapt to real-time feedback.

In addition to task-based classification, social navigation can be categorized by communication strategies, focusing on how robots interact with humans through signals. For a more in-depth discussion on taxonomy, refer to Singamaneni et al. (2024) and Mirsky et al. (2021).

This review focuses on independent (crowd-aware) navigation due to its broad applicability. Its core principles can be extended to assistive and collaborative tasks, making it a more general foundation for various social navigation tasks.

Imitation learning (IL) enables learning an end-to-end policy directly from expert demonstrations, bypassing the need for hand-crafted rewards. While inverse reinforcement learning (IRL) infers a reward function from human demonstrations or pedestrian datasets (Kim and Pineau, 2016; Fahad et al., 2018) then learns a policy, behavioral cloning (BC) learns actions directly from demonstrations but struggles in dynamic settings due to its reliance on fixed data. More advanced IL approaches aim to overcome these limitations. One of the earliest data-driven approaches for static obstacle navigation, proposed by Pfeiffer et al. (2017), uses a goal-conditioned model with 1D CNN and pooling layers trained using BC. The model takes in 2D LiDAR readings and goal information to predict actions and is trained on demonstration data collected using the dynamic window approach (DWA) planner (Fox et al., 1997). While effective in static environments, this approach does not incorporate past observations, reducing its effectiveness in dynamic obstacle scenarios. Similarly, CANet (Long et al., 2017) applies behavioral cloning to learn a navigation policy from multi-agent data generated using ORCA planner (Van Den Berg et al., 2011). The model is an MLP trained to output a probability distribution over 61 pre-defined 2D velocity clusters, capturing a range of socially aware navigational behaviors. A value iteration network (VIN)-based planner, proposed by Liu et al. (2018), applies VIN (Tamar et al., 2016) to social navigation. VIN introduces a neural network architecture with a differentiable planning module that approximates the classical value iteration algorithm. Given a reward map and local transition model, VIN iteratively maps rewards and previous value estimates into Q-values using convolutional layers, where each channel corresponds to an action’s outcome. A channel-wise max pooling layer retrieves the maximum over actions, yielding the updated value function, which is then used by a greedy reactive policy network (e.g., softmax) to generate an action distribution. Liu et al. (2018) extend VIN by adding an MLP that combines the VIN output with the robot’s velocity to predict actions. Trained in a supervised manner on real and synthetic maps with demonstration actions derived from a reactive optimization-based planner, this approach provides a novel perspective on navigation. However, it’s limited to static environments and needs to be extended to dynamic settings with crowds. Another approach, GAIL, is used by Tai et al. (2018) to train a navigation policy. GAIL employs a generator (policy) that processes depth images to predict actions, while a discriminator distinguishes between the generator’s actions and expert demonstrations. To stabilize training, the discriminator is defined as a regression network, inspired by WGAN (Arjovsky et al., 2017), rather than a standard classifier. Initially, the policy is pre-trained with behavioral cloning on expert data and then fine-tuned using TRPO with the discriminator. The main advantage of GAIL is its use of online simulation-based training, which helps mitigate generalization issues. MuSoHu (Nguyen et al., 2023) addresses data scarcity in data-driven navigation by providing a large-scale dataset of 100 km of human navigation patterns collected with a helmet-mounted sensor suite. Applying behavioral cloning on this dataset produces a human-like path-planning policy that mitigates behavior modeling inaccuracies and shows strong real-world performance. DeepMoTIon (Hamandi et al., 2019) aims to mimic human pedestrian behavior by using imitation learning to train a navigation policy. The approach uses pedestrian datasets to simulate human-centric LiDAR data, training an LSTM-based policy through supervised learning. The model predicts the pedestrian’s future direction and velocity based on its LiDAR data and final goal. To account for variability in human behavior, it employs a Gaussian distribution for direction prediction, enabling the capture of diverse movement patterns in similar scenarios.

World models provide agents with internal representations of environment dynamics, enabling more informed, end-to-end decision-making. One prominent example is NavRep (Dugas et al., 2021), which integrates the World Model framework (Ha and Schmidhuber, 2018) with the PPO algorithm to train a policy. NavRep introduces rings, a novel 2D LiDAR representation that arranges data into exponentially spaced radial intervals within a polar coordinate grid, enhancing close-range resolution. Similarly, Cui et al. (2021) applies world models with the TD3 algorithm in a MARL framework, with the state represented by stacked 2D obstacle maps generated from multiple LiDAR scans.

Most methods discussed so far rely on a single sensor input, which can be prone to noise and limited in accuracy. To enhance perception robustness for end-to-end systems, sensor fusion techniques are employed. For example, Liang et al. (2021) processes 2D LiDAR data using a 1D CNN and depth images using a 2D CNN, with inputs collected over three consecutive time steps, and combines the outputs through concatenation. In another approach, Han et al. (Han Y. et al., 2022) propose a fusion network that integrates RGB images and 2D LiDAR data to produce depth information. The 2D LiDAR data is first transformed into the camera’s coordinate frame, then combined with RGB data through an encoder-decoder CNN network (Ma and Karaman, 2018) to produce a depth image. The depth image is processed by a self-attention module, which prioritizes pixels based on factors such as robot type, goal position, and velocity, thus enhancing the agent’s situational awareness. Some navigation systems focus on optimizing performance with sensors that have limited fields of view. In these setups, self-supervised and supervised approaches are used to improve the agent’s situational awareness. For example, Choi et al. (2019) employ an actor-critic algorithm where the actor network uses an LSTM, while the critic receives additional information, such as a local 2D map. This approach allows the actor to rely on temporal cues, while the critic aids in evaluating action choices more accurately. Similarly, Monaci et al. (2022) introduce a method where an initial policy is trained using privileged information, such as precise human positions within the environment. This policy is subsequently distilled into a non-privileged policy that learns to approximate the privileged information through supervised learning.

Multi-objective reinforcement learning (MORL) (Roijers et al., 2013) frameworks are increasingly applied in end-to-end navigation tasks where agents must balance multiple, often conflicting, objectives. MORL allows a policy to be trained on multiple different objectives, enabling the adjustment of objective weightings, referred to as a preference vector, during deployment (Hayes et al., 2022). This flexibility is particularly beneficial in dynamic social environments, where safety, efficiency, and comfort are key yet sometimes competing. For example, Cheng et al. (2023) implement a vectorized Q-learning-based MORL algorithm to train a policy with a simulated crowd. Meanwhile, Choi et al. (2020) use the SAC MORL algorithm to train a navigation policy with a preference vector learned from human feedback, sampled through a Bayesian neural network (Blundell et al., 2015). Hierarchical reinforcement learning (HRL) divides complex tasks into manageable sub-tasks or sub-goals, allowing an agent to focus on different levels of decision-making. In HRL architectures, the high-level policy selects sub-goals, while the low-level policies execute these sub-goals through specific navigation actions. For instance, Lee et al. (2023) propose an HRL framework in which the high-level policy focuses on reaching the goal efficiently, minimizing time-to-goal. This policy generates a skill vector, which is then interpreted by the low-level policy to execute specific navigation skills, such as collision avoidance, goal-reaching, and maintaining a safe distance. Both levels of policy utilize 2D LiDAR data and goal state information. Other HRL approaches offer variations in task distribution and shared information. Zhu and Hayashibe (2022) use a high-level policy as a safety controller to halt the low-level policy if necessary, while Wang et al. (2021) implement an HRL framework in which the high-level policy shares a sub-goal with the low-level navigation policy.

In vision-based end-to-end navigation, RGB or RGB-D cameras provide input for agents to reach goals specified by relative position (PointGoal), target images (ImageGoal), or instructions (Vision-Language Navigation). These planners excel in visually rich settings without global maps, relying solely on relative goal information. Policies typically use CNN-RNN architectures, where CNNs process images and RNNs build an internal map (Kulhánek et al., 2019). Even blind agents, lacking vision but using memory-based policies, can navigate efficiently via spatial awareness and wall-following strategies (Wijmans et al., 2023). Such methods use photorealistic simulators based on real-world scans (Chang et al., 2017) and often employ discrete actions for training efficiency. Vision-based social navigation is emerging, with proximity-aware (Cancelli et al., 2023) and Falcon (Gong et al., 2024) methods using auxiliary tasks to better anticipate and navigate around pedestrians and obstacles.

Vision-language models (VLMs) are powerful multimodal models with the ability to support navigation through reasoning, visual grounding, and contextual understanding. Early work on vision-language navigation (VLN) (Anderson et al., 2018a) introduced text-based high-level planning, which can be extended to social navigation for local decision-making (Li et al., 2024). Beyond high-level planning, several recent hybrid methods integrate VLMs directly into the social navigation pipeline. Song et al. (2024) use a VLM to select high-level direction and speed, which are integrated with goal and obstacle costs in a model-based planner, with weights determined through an additional VLM prompt. GSON (Luo et al., 2025) leverages VLMs to detect social groups and integrates the results into an MPC planner to generate paths that avoid them. OLiVia-Nav (Narasimhan et al., 2025) distills social context from a large VLM into lightweight encoders that provide semantic inputs to a trajectory planner, which then generates candidate motions and selects the one most aligned with captions distilled from expert demonstrations. OLiVia-Nav further incorporates lifelong learning to update its encoders with new data. Related to this, Okunevich et al. (2025) introduce an online learning approach that adapts a social module in real time, updating the social cost function during deployment. Alternatively, coding-capable large language models (LLMs) have been prompted to generate reward functions from natural language preference descriptions (Ma et al., 2023), with applications in navigation and preference alignment (Wang et al., 2024). Social-LLaVA (Payandeh et al., 2024) leverages a VLM fine-tuned for social robot navigation to directly map decisions onto a predefined set of low-level navigation primitives. Despite this progress, the slow inference and high computational demands of VLMs currently limit their use for real-time reactive social navigation. As a result, they are mostly applied as global planners, semantic encoders, or social-context modules, while their broader potential remains underexplored.

Beyond RL, self-supervised methods enable partial or full training of navigation policies using generated labels. For example, Hoeller et al. (2021) train a VAE to encode depth data, filter noise, and enhance sim-to-real transfer, providing informative representations for faster RL training. Yang et al. (2023) propose a bi-level framework where a neural network predicts waypoints optimized through a differentiable ESDF-based cost function; while deployment is simplified by using a spline to fit waypoints. Roth et al. (2024) further incorporate semantic costmaps, though dynamic obstacle avoidance remains unevaluated.

Overall, end-to-end navigation directly maps sensor inputs to actions and supports continuous actions, multi-agent training, model-based RL, multi-objective and hierarchical frameworks, VLMs, and self-supervised learning. However, challenges remain in ensuring safety and robustness.

Approaches that incorporate human demonstrations and preferences into policy training have proven effective for aligning robot behavior with human expectations in social navigation. De Heuvel et al. (2022) use the SAC algorithm with behavioral cloning to train a policy in simulation, closely fitting human demonstration trajectories collected via a VR pointer. This work is extended in De Heuvel et al. (2023) by adding a perception pipeline that predicts future human positions. Building on this, De Heuvel et al. (2024) employ MORL-TD3 with multiple objectives, including a human demonstration distilled into a reward function using D-REX. Lastly, Marta et al. (2023) adopt a multi-objective approach to balance an expert-designed objective with a human preference objective derived from a reward model trained on pairwise human trajectory comparisons.

Overall, human position-based navigation utilizes explicit knowledge of human positions and velocities to enable safer and more socially-aware navigation policies. Techniques such as CADRL-based methods establish foundational frameworks by learning interaction-aware value functions. Moreover, incorporating human preferences and demonstrations ensures policies align closely with human expectations.

GazeNav (Chen et al., 2020a) employs a model-based RL approach with gaze-based attention that uses 2 two-layer Graph Convolutional Networks (GCNs) to define its value function. The first GCN, an attention network, treats the robot and humans as graph nodes with uniform edge weights, predicting attention weights for each connection. The second GCN is an aggregation network that uses the predicted attention weights as edge values to compute embedding vectors for each human-robot pair, which are then passed into an MLP-based value function. To train the attention network, GazeNav introduces three supervised methods: uniform weights, distance-based weights, and gaze-modulated weights. The gaze-modulated weights are obtained by tracking human gaze in a simulated environment, assigning higher attention to humans within the gaze direction. Experiments show that gaze-modulated weights outperform uniform, distance-based, and self-attention-based weights (Chen C. et al., 2019), demonstrating the benefits of incorporating human gaze data. For a more expressive representation, Navistar (Wang W. et al., 2023) uses a three-block architecture to model spatio-temporal crowd interactions. A spatial block (GCN plus multi-head attention) creates spatial embeddings; a temporal block applies multi-head attention with positional encoding for each human. A multi-modal transformer block then merges these outputs using cross-attention and self-attention to produce the final action and value outputs. In a related approach, Liu Z. et al. (2023) integrate GNNs with occupancy grids to capture spatial-temporal characteristics. At each time step, the environment is divided into a robot-centered grid and an obstacle-centered grid for each human, both processed by a CNN. The CNN outputs are then passed through an LSTM to capture temporal patterns, feeding into a Graph Attention Network (GAT) that produces interaction-aware embeddings. The control policy uses an MLP to generate action distributions from the GAT’s aggregated output.

To summarize, human attention-based navigation methods explicitly model human-human and human-robot attentions to enable socially-aware and adaptive policies. Approaches utilizing pooling layers, GNNs and RNNs, provide improved social compliance by capturing spatial and temporal relationships.

MCTS-RNN (Eiffert et al., 2020a) is a model-based RL navigation system that uses an LSTM encoder-decoder human prediction model as its dynamics model. The LSTM model is trained on pedestrian datasets and outputs a Gaussian distribution over future human states. Planning is conducted using MCTS with a receding horizon, performing single-step rollouts from each node to reduce runtime, which increases state uncertainty. To handle this, the reward function includes both goal proximity and prediction uncertainty. MP-RGL (Chen C. et al., 2020) integrates MCTS planning with a GCN-based human prediction model. The GCN operates on a fully connected graph comprising humans and the robot, where edge weights are computed using Gaussian similarity in the node embedding space (Wang X. et al., 2018). Planning is performed through a simplified MCTS (Oh et al., 2017), with a d -step planning horizon, leveraging value function estimates instead of explicit rollouts.

GO-MPC (Brito et al., 2021) is a hybrid framework that integrates RL and nonlinear MPC for navigation, where an LSTM-based RL model proposes sub-goals (as Gaussians) and the MPC computes optimal, collision-free trajectories to these sub-goals. The RL model is first supervised-trained with MPC-generated labels, then fine-tuned with PPO, aiming to maximize goal-reaching and minimize collisions. The MPC minimizes distance and control costs, enforcing constraints to avoid predicted human paths. Poddar et al. (2023) propose a hybrid approach that integrates a Social-GAN (Gupta et al., 2018) human prediction model with an MPC planner. This approach uses discrete MPC to optimize a cost function that balances goal distance, social distance, and alignment with Social-GAN predictions to encourage human-like behavior. While Social-GAN can generate multiple predictions per human, results indicate that single and multiple prediction scenarios perform comparably to simpler constant-velocity estimates.

SARL-SGAN-KCE (Li et al., 2020) combines Social-GAN predictions with the SARL model (Chen C. et al., 2019) to choose optimal single-step actions. To ensure smooth motion, the planner constrains the action space by limiting angular velocity and penalizing rapid acceleration changes. Experimental results show that a higher number of trajectory predictions per human achieves performance comparable to a lower number of predictions. Finally, Liu S. et al. (2023) propose a model-free PPO RL approach that incorporates off-the-shelf human prediction models like GST (Huang et al., 2021). Human predictions are processed with multi-head human-human attention, then through robot-human attention with the robot’s state, followed by a GRU that outputs the value and action. The reward penalizes intersecting predicted human paths, reducing collision risk despite prediction uncertainty.

In summary, human prediction-based navigation enhances decision-making by anticipating future human movements, enabling more strategic and socially compliant planning. Challenges include managing uncertainty from the robot’s impact on human behavior and the computational cost of tree-based methods like MCTS, which require repeated action sampling and forward simulation.

Hybrid multi-policy planning combines multiple strategies, where robots switch policies based on context and uncertainty. For example, Sun et al. (2019) switches between RL and RVO when a collision is imminent. Katyal et al. (2020) build on this with risk-averse and aggressive policies. By default, the system follows the aggressive policy but switches to the risk-averse policy in novel social scenarios, identified by an LSTM-based probabilistic pedestrian prediction module that uses goal intent prediction to generate a set of possible trajectories. The policy selector computes uncertainty from these predictions, with higher uncertainty indicating unfamiliar situations where the risk-averse policy is preferred. Extending this approach, Fan et al. (2020) develop a three-policy system with a scenario classifier to switch between a PID controller, a standard RL policy (Long et al., 2018), and a safe RL policy with clipped velocity. The classifier relies on two parameters, the safe radius and risk radius, based on the distance to nearby obstacles. When within the safe radius, the PID policy is used. In the risk radius, the RL policy takes over, and outside both, the safe policy is employed. To address more complex scenarios, Amano and Kato (Amano and Kato, 2022) add a fourth policy to this setup, a reset policy to move the robot toward a larger unoccupied space if it detects a freezing robot scenario. This extension ensures the robot can navigate out of potentially freezing situations. Furthermore, Linh et al. (2022) propose a multi-policy system with three policies, using an RL-based policy selector to choose the most appropriate policy dynamically. Policies include both learning-based (RL) and model-based (TEB) planners (Rösmann et al., 2015). The selector is trained to optimize rewards by picking the best policy for a given context, combining flexibility with performance for complex navigation tasks.

Nishimura and Yonetani (2020) introduce Learning-to-Balance (L2B), a single-policy RL system that dynamically switches between two behaviors: passive crowd avoidance or active path-clearing through audible signals. The robot action is defined by a velocity vector and a binary mode indicator, with a reward function that discourages excessive path-clearing while promoting social distancing. To simulate the impact of path-clearing sounds on human behavior during training, L2B uses a simplified version of emotional reciprocal velocity obstacles (ERVO) (Xu M. et al., 2019), which accounts for emotional reactions to perceived threats. IAN (Dugas et al., 2020) is a multi-policy navigation system that uses Monte Carlo Tree Search (MCTS) to choose among three planning policies: intend (RVO planner (Alonso-Mora et al., 2013) for reactive avoidance), say (verbal path announcement with lower speed and assumed human cooperation), and nudge (DWA planner (Fox et al., 1997) for cautious progress). MCTS evaluates paths by crowdedness, perceptivity, and permissivity, selecting the lowest-cost route and adapting plans based on each policy’s success probability. Both L2B and IAN require the robot to have a speaker and operate where its audio signals are audible.

Lütjens et al. (2019) propose a hybrid safe RL system based on discrete MPC, optimizing a cost function that accounts for estimated goal-reaching time and predicted collision probability. An ensemble of LSTMs predicts collision probabilities of motion primitives, with MC-dropout (Gal and Ghahramani, 2016) used for uncertainty estimation. The collision prediction model is trained as a binary classifier in simulation, penalizing uncertainty to encourage safe exploration. However, this approach heavily depends on collision model accuracy, and inaccuracies can lead to overly conservative behavior. Sathyamoorthy et al. (2020b) introduce Frozone, which prevents robot freezing by detecting potential freezing zones (PFZs) using pedestrian positions and velocities. A convex hull is constructed around predicted pedestrian locations, and the robot computes a deviation angle to avoid these regions. However, in confined spaces like corridors, Frozone may lead the robot toward other obstacles. XAI-N (Roth et al., 2021) leverages decision trees to create an interpretable navigation policy. XAI-N distills an RL policy (Fan et al., 2018) into a single decision tree using the VIPER method (Bastani et al., 2018), prioritizing modifiability and transparency over continuous action control. To enhance performance, the approach incorporates decision rules to address safety challenges such as freezing and oscillation, making it a more reliable option for social navigation.

Bansal et al. (2020) propose a Hamilton–Jacobi reachability-based framework that augments the human state with a belief over future intent, producing a forward reachable set that includes all likely pedestrian states for fixed time-horizon with probability above threshold λ . These sets are incorporated as time-dependent obstacles and avoided with a spline-based trajectory planner. However, in highly populated scenes, the predicted reachable sets may overlap heavily for a low threshold, effectively blocking all routes and causing the robot to freeze, and while the human model parameters can be learned from data, the overall prediction approach is model-based. Jang and Ghaffari (2024) learn social zones from pedestrian data by relating distance to line-of-sight angle, then approximate them with speed-dependent ellipses enforced through a control-barrier function within a hybrid MPC planner. The learned zones are front-biased—indicating humans prefer more space ahead—and slightly tilted, reflecting overtaking behavior. While enforcing them as hard constraints improves safety, it can be overly conservative, since people typically tolerate reduced spacing in crowded settings. CASRL (Zhou et al., 2023) frames safety in navigation as a multi-task RL problem: goal reaching and collision avoidance tasks. It extends an off-policy RL algorithm (TD3) with separate critics for each task, while the actor is updated using a conflict-averse rule that maximizes the minimum improvement across tasks. This reduces performance loss when gradient updates disagree. However, the reported simulation gains fall short and may require further tuning.

Constrained RL provides a natural framework for enforcing safety, as constraints take precedence over the reward objective when violated. For instance, Pfeiffer et al. (2018) introduce a safe RL navigation policy that defines a collision constraint, trained using constrained policy optimization (CPO) (Achiam et al., 2017), which maximizes reward while constraining the expected number of collisions. SoNIC (Yao et al., 2024) introduces a safety constraint derived from Adaptive Conformal Inference (ACI), which quantifies the uncertainty of predicted pedestrian trajectories. Similarly, Zhu et al. (2025) propose a confidence-weighted trajectory prediction model, where a Bayesian β parameter adapts the uncertainty region based on prediction errors. In their method, uncertainty is incorporated through a robust dynamical distance constraint that estimates time-to-collision, rather than relying on simple distance-based thresholds. However, both approaches employ trajectory predictors that neglect the robot’s presence, resulting in predictions where pedestrians are assumed to move independently of the robot. This leads to overly conservative robot behavior.

In conclusion, safety-aware navigation improves reliability in learning-based systems through structured mechanisms, but further work is needed to balance safety with efficiency and ensure adaptability to diverse real-world scenarios.

The reward function for reaching a goal state is the main component of any navigation task. It is often defined as an indicator function [ ‖ p r − g r ‖ ≤ g tol ] , which returns 1 when the Euclidean distance between the robot’s current position p and the goal position g r is within a specified tolerance g tol . A dense reward formulation for goal-reaching provides feedback based on progress, calculated as ‖ p t r − g ‖ − ‖ p t − 1 r − g r ‖ , reflecting incremental advancement toward the goal (Long et al., 2018; Tan et al., 2020). Other approaches reward the agent for moving along grid cells that align with the global path (Liu Z. et al., 2023).

The reward function for collision avoidance is often defined as an indicator function, − [ p r ∈ C ] , which returns 1 when the robot’s position p r is within a collision set C , or when a collision is detected using other means. Alternative definitions penalize proximity to obstacles, such as − [ d obs ∣ d obs ≤ d tol ] (Chen et al., 2017b) or [ d obs d tol − 1 ∣ d obs ≤ d tol ] (Cui et al., 2021) where d obs is the minimum distance to nearby obstacles and d tol is the tolerance distance. These formulations may apply different thresholds for static and dynamic obstacles, such as humans (Cui et al., 2021). Other approaches use step functions to gradually increase the penalty as the robot gets closer to obstacles, encouraging safer navigation.

To encourage efficient and timely navigation, reward functions often include terms that promote higher speeds. This may take the form of a gradual step function that provides a higher reward for increased velocity (Lee and Jeong, 2023) or a negative step-cost applied at each timestep to minimize time taken to reach the goal (Wang Y. et al., 2018; Choi et al., 2019).

For smooth trajectory generation, a negative reward proportional to the rotational velocity, − | ω | is applied for differential drive robots (Tan et al., 2020; Xie and Dames, 2023). Additionally, Hoeller et al. (2021) penalizes lateral and backward velocities by adding a negative reward proportional to their squared magnitudes, encouraging smoother and more consistent forward motion.

Social norms can be integrated into the reward function to promote behaviors like passing, crossing, and overtaking in socially appropriate ways (Chen et al., 2017c). This reward function is typically defined as a conditional function based on human parameters relative to the robot, including x-axis position, velocity, distance, relative heading angle, and heading angle difference. For instance, to promote overtaking from the left, the robot is rewarded when certain conditions are met: the goal distance exceeds 3, the human is positioned within 0 < p x < 3 , and 0 < p y < 1 , the robot’s velocity surpasses the human’s, and their heading angle difference is under π / 4 .

Model-based or geometric rewards using human position have enabled more robust navigation. For instance, DRL-VO (Xie and Dames, 2023) uses velocity obstacles (VOs) to model human motion, rewarding alignment with the optimal heading direction, where VOs are computed during training only. Han R. et al. (2022) incorporate VOs into both state and reward, with rewards based on joint VO area, velocity differences, and estimated minimum time to collision. Zhu et al. (2022) employ an oriented bounding capsule (OBC) model, where human velocity adds a buffer in front of the OBC, and the reward is the minimum distance to the OBC; OBC parameters are included in the robot’s state for better learning. Additionally, Samsani and Muhammad (2021) define a danger zone (DZ) as an extended sector around humans, accounting for uncertainty in position and velocity predictions.

Reinforcement Learning from Human Feedback (RLHF) provides a framework to simultaneously learn a policy and a reward function using human input (Christiano et al., 2017), with applications spanning various domains, including language models like GPT-3 (Ouyang et al., 2022). In social navigation, Wang R. et al. (2022) applies RLHF to learn a reward function based on pairwise human preferences over trajectory segments.

For planners that utilize human trajectory predictions, a negative reward is often used to discourage the robot from intruding into human-predicted zones (Liu S. et al., 2023). Additionally, a negative reward can be defined over prediction uncertainty, as in (Eiffert et al., 2020b), where the reward is the negative square root of the determinant of the covariance matrix, d e t ( Σ ) , encouraging actions that lead to more predictable crowd behavior.

Exploration rewards are designed to encourage the robot to explore a wide range of actions or states. Action-based exploration rewards promote action diversity by maximizing the policy’s entropy (Schulman et al., 2017), while state-based exploration rewards encourage the robot to explore new areas. For instance, the intrinsic curiosity module (ICM) (Pathak et al., 2017), applied to navigation tasks (Shi H. et al., 2019; Martinez-Baselga et al., 2023) to reward the robot for discovering novel states, thereby enhancing its learning process.

Task-specific rewards are custom-designed to achieve the requirements of a particular navigation task. For example, in social navigation with a human companion, Li et al. (2018) define a distance-based reward that penalizes the robot for straying from its companion, encouraging it to stay close and coordinate its movement with the human partner.

Inverse reinforcement learning (IRL) infers a reward function from expert demonstrations, either by using handcrafted state–action features (Okal and Arras, 2016; Kim and Pineau, 2016) or by learning feature representations directly with neural networks (Fahad et al., 2018). For instance, Vasquez et al. (2014) learn a reward function expressed as a weighted combination of features that capture local crowd density, relative velocities and orientations of nearby pedestrians, the robot’s own velocity, and social force interactions. More recently, methods like disturbance-based reward extrapolation (D-REX) (Brown et al., 2020) learn reward functions from suboptimal or unlabeled data. D-REX applies behavioral cloning, adds increasing ϵ -greedy noise, and trains the reward to favor less noisy trajectories, promoting desirable behavior, an approach effective for social navigation (De Heuvel et al., 2024). Similar techniques include T-REX (Brown et al., 2019) and SSRR (Chen et al., 2021).

Various techniques have been developed to automatically determine the optimal values of each objective weight w i , eliminating the need for manual tuning. One such method is inverse reinforcement learning (IRL), which infers the weights w i by matching the robot’s behavior to expert demonstrations (Ziebart et al., 2008). Additionally, AutoRL (Chiang et al., 2019; Parker-Holder et al., 2022) employs automated hyperparameter tuning to optimize the reward weights w i during training to enhance task-specific performance metrics.

Crowd datasets play a critical role in advancing data-driven approaches for both crowd behavior simulation and human trajectory prediction. They provide the necessary information to model realistic crowd interactions and dynamics, as detailed in the Appendix. Additionally, these datasets support the training of human prediction methods, as explored in Section 3.3.2. Table 7 organizes these datasets by their sensory platforms, including stationary sensors, moving robots, and moving vehicles, each serving distinct purposes and applications. While long-term crowd tracking datasets such as the ATC dataset (Brščić et al., 2013) exist, they lack the scale and diversity needed to support social navigation research.

Simulators provide a controlled virtual environment for developing and evaluating social navigation algorithms by modeling human-robot interactions and crowd behaviors. Table 8 categorizes simulators based on key attributes, such as the supported sensor types, the human model ranging from simple cylindrical shapes to detailed 3D figures, supported crowd behaviors, evaluation metrics based on implementation specifics.

Many simulators share a common emphasis on creating realistic environments. For instance, Habitat (Szot et al., 2021) and Gibson (Li et al., 2021), widely used in embodied AI research, render highly detailed indoor spaces using real 3D scans. However, these environments are typically limited to smaller areas like apartments or offices, making them less suitable for large-scale crowd simulations. Additionally, several simulators, such as Isaac Sim (Makoviychuk et al., 2021), prioritize achieving high FPS, which is crucial for training performance. Simulators also vary in the complexity of crowd behaviors, with some supporting basic movement patterns and others providing sophisticated, behavior-rich models that more accurately capture crowd dynamics, such as NavRep (Dugas et al., 2021). However, current simulators remain limited, as an efficient RL-supported crowd simulation with diverse scenarios is still missing, which we aim to address with our benchmark.

Human detection methods are typically tailored to specific sensors, including 2D LiDAR, 3D LiDAR, RGB, and RGB-D sensors. Tracking enhances detection by assigning unique identifiers to individuals and addressing challenges such as sensor occlusions, which is essential for reliable multi-object tracking (MOT). While human detection provides the robot-centric positions of each detected person, tracking maintains a history of these positions over time, enabling the estimation of their velocities.

Predicting human trajectories is critical for effective social navigation. Traditionally relying on knowledge-based methods, the field has shifted towards learning-based approaches, which consistently outperform traditional methods on metrics such as average displacement error (ADE) (Pellegrini et al., 2009) and final displacement error (FDE) (Alahi et al., 2016). Learning-based methods leverage crowd datasets (see Section 3.2.1) and typically employ CNN, LSTM, or GAN architectures (Korbmacher and Tordeux, 2022).

Pre-training techniques, such as behavioral cloning from demonstrations (Pfeiffer et al., 2018; Chen C. et al., 2019), accelerate training by providing basic navigation skills and reducing RL exploration. Self-supervised methods, like VAEs with reconstruction loss (Dugas et al., 2021; Hoeller et al., 2021), improve state representation, while transfer learning from pretrained CNNs enhances RGB input processing (Hong et al., 2021). Policy transfer from existing models is also used (Wijmans et al., 2019). These approaches improve training efficiency, convergence, and generalization.

Auxiliary tasks are additional tasks or objectives incorporated during training to support learning the main task. This offers better training signal and model performance. Auxiliary tasks have been shown to improve navigation performance by training models to predict features such as depth, loop closures (Mirowski et al., 2016), and location estimation (Tongloy et al., 2017). Additional tasks include predicting immediate reward prediction and learning to control specific regions in the input image (Jaderberg et al., 2016) or predicting image segmentation (Kulhánek et al., 2019). In social navigation, auxiliary tasks are used to improve understanding of social dynamics. For instance, Proximity-Aware (Cancelli et al., 2023) incorporates tasks to estimate the distance and direction of surrounding humans, while Falcon (Gong et al., 2024) incorporates tasks for predicting the number of nearby humans, tracking their locations, and estimating their future trajectories. These tasks enable the model to acquire valuable insights into the environment’s social dynamics, leading to more efficient and informed planning.

Curriculum learning gradually increases task difficulty during training, aiding convergence in challenging social navigation tasks. In RL, this process involves three steps: task generation, sequencing, and transfer learning (Narvekar et al., 2020). Task generation creates scenarios of varying difficulty by adjusting obstacles, goal distances, or map complexity, using parameter sampling or grid search. Sequencing organizes tasks by increasing difficulty, either at a fixed rate or adaptively based on agent performance, and may involve modifying reward functions or start/goal distributions (Riedmiller et al., 2018; Florensa et al., 2018), optimization strategies (Matiisen et al., 2019), Curriculum MDPs (Narvekar et al., 2017), or human feedback (Bengio et al., 2009). Transfer learning adapts agents when intermediate tasks differ in state/action spaces, rewards, or dynamics, such as transitioning from precise states to noisy sensors, or from indoor to outdoor navigation. This combination allows agents to efficiently learn complex social navigation skills.

The teacher-student framework enables a teacher model, often trained with privileged information, to guide a student via real-time feedback, reward shaping, or action labels. Knowledge transfer is achieved through policy distillation (Rusu et al., 2015), using labeled paths or actions from the teacher, student, or both (Czarnecki et al., 2018), allowing the student to imitate and refine its navigation policy, which can later be fine-tuned with RL. Teachers may also provide reward signals to enhance exploration (Czarnecki et al., 2019) and corrective action feedback (Ross et al., 2011). Model-based teachers like MPC are also used (Lowrey et al., 2018). Asymmetric actor-critic methods allow the critic to use privileged information to guide the actor (Pinto et al., 2017). In teacher-student curriculum learning, teachers assign progressively harder tasks and are rewarded for student improvement (Matiisen et al., 2019), while multi-teacher approaches combine skills from specialized teachers (Rusu et al., 2015). For social navigation, non-optimal teachers (e.g., PID planners) can be combined with RL, accelerating training by switching to the higher Q-value source (Xie et al., 2018).

Sim-to-real transfer for navigation tackles the challenge of adapting a simulation-trained policy to perform reliably in real-world environments. Achieving sim-to-real transfer requires a highly realistic simulator (refer to Section 3.2.2) and the implementation of techniques like domain randomization and domain adaptation. These techniques operate at different levels: scenario-level randomization and adaptation (see Appendix for details) modify various aspects of the simulated environment, while sensor-level noise enables the policy to handle discrepancies in real-world sensor data. Domain adaptation adjusts simulation-trained models to real-world domains. For RGB data, this uses real-world samples and methods like discrepancy minimization, adversarial alignment, or reconstruction methods for feature alignment (Wang and Deng, 2018). For depth sensors, techniques such as depth completion and refinement address real-world limitations, improving consistency with simulated data (Khan et al., 2022). Domain randomization narrows the sim-to-real gap by introducing simulated variability, allowing policies to generalize to real-world conditions (Tobin et al., 2017). For RGB inputs, this includes varying visual features to simulate lighting and color changes (Anderson et al., 2021); for depth sensors, it involves adding noise, occlusions, warping, and quantization (Muratore et al., 2022; Thalhammer et al., 2019). Active domain randomization further improves robustness by focusing on model-effecting variations (Mehta et al., 2020; Zakharov et al., 2019).

Evaluating social navigation policies in real-world settings is crucial for assessing their robustness, adaptability, and social acceptability. Experiments typically fall into three categories: experimental demonstrations, lab studies, and field studies (Mavrogiannis et al., 2023). Experimental demonstrations offer proof-of-concept with limited reproducibility (Chen et al., 2017b; Chen C. et al., 2019), while lab studies provide structured, repeatable tests in controlled environments with systematic reporting (Tsai and Oh, 2020; Mavrogiannis et al., 2019). Field studies are the most comprehensive, deploying robots in public spaces among uninstructed pedestrians (Kato et al., 2015; Kim and Pineau, 2016). Real-world evaluations combine quantitative metrics with qualitative observations, such as participant feedback or questionnaires, to assess social adaptability and compliance (Pirk et al., 2022).

Navigation and social navigation metrics provide a structured framework to assess robot performance in crowded environments. Traditional navigation metrics assess robots’ fundamental abilities such as reaching targets and avoiding obstacles, while social navigation metrics focus on interactions with humans, including maintaining personal space and minimizing disruptions to bystanders. Together these metrics, as detailed in Table 9, guide the development of navigation systems that achieve task objectives efficiently while adhering to socially appropriate behaviors, promoting safer and widely accepted robot deployments.

The baseline planners include ORCA (Van den Berg et al., 2008), the SFM (Helbing and Molnar, 1995), and DWA (Fox et al., 1997). They serve as classical foundations for comparison with advanced methods. Each is given privileged access to the map layout and all human positions, ensuring optimal performance under ideal conditions.

The end-to-end planner is based on the RL policy from Fan et al. (2020), which processes recent 2D LiDAR scans with a 1D CNN, combines them with the robot’s state, and uses an MLP for action selection. Due to its suboptimal performance, we adopt an RNN-enhanced architecture (Hoeller et al., 2021), where the CNN output and robot state are fed into a GRU network, improving results. This end-to-end model learns navigation directly from sensor data, without using human state information.

We implement Behavioral Cloning (BC) for imitation learning, offering a simple alternative to methods like GAIL (Ho and Ermon, 2016) without needing simulated environments. Trained on 35,000 successful human attention-based planner episodes, matching the planner’s performance would signal robust generalization from real-world data. The network architecture mirrors the human attention-based planner.

The SARL planner (Chen C. et al., 2019) employs an attention-based network to model robot-human attentions. We extend the original value network to an actor-critic framework and add a LiDAR network (Fan et al., 2020) for static obstacle handling. Unlike Liu et al. (Liu L. et al., 2020), which switches between separate policies for human and non-human scenarios, our approach uses a learned embedding to pad human input when no humans are present.

The prediction-based planner adapts the RGL model (Chen C. et al., 2020), integrating robot state and LiDAR input to predict human trajectories in the robot frame. These predicted trajectories are processed by an actor-critic policy, following the SARL planner (Chen C. et al., 2019), to handle fixed-size trajectories. When no humans are present, a learned embedding pads the input for consistency.

Inspired by Linh et al. (2022), the safety-aware planner combines ORCA (Van den Berg et al., 2008) for static environments and the human attention-based planner for dynamic settings, using a policy switcher based on obstacle proximity. This hybrid approach balances safety and efficiency by adapting to both static and human-dense scenarios.

Most safety-oriented planners, like multi-policy approaches (Sun et al., 2019; Katyal et al., 2020; Fan et al., 2020), assume that reducing speed enhances safety, but this is not always valid; rapid maneuvers may be needed in dynamic, crowded settings. Relying solely on speed reduction can compromise safety in complex environments. Similarly, several human-prediction-based methods (Yao et al., 2024; Zhu et al., 2025) primarily forecast human motion without explicitly modeling the robot’s influence on the crowd, which limits their ability to generate safe and adaptive plans. Instead, planners should learn context-aware safe behaviors, adjusting speed as needed and responding to emergencies, to achieve both safety and efficiency without unnecessary conservativeness.

A major limitation to model generalizability is the limited diversity of training scenarios. Future benchmarks should incorporate a wider range of realistic, data-driven scenarios reflecting true pedestrian distributions and start-goal configurations. Long-term crowd tracking datasets similar to, or even larger than ATC (Brščić et al., 2013), which capture varied environments in a shopping mall, can help provide such diversity.

Physics simulators range from simple kinematic to detailed dynamic models, with high-fidelity simulation improving sim-to-real transfer and enabling robot-specific planners that can integrate low-level control, such as direct wheel velocities. Likewise, accurate sensor simulation enhances robustness; while most simulators use generic models for simplicity (Inc D, 2025), sensor-specific models that replicate real-world parameters and noise can significantly improve generalization and sim-to-real performance.

Most crowd simulation methods focus on human-human and human-obstacle interactions, but accurately modeling human-robot interactions remains a challenge. Some approaches ignore the robot’s presence (Chen Y. et al., 2020; Dugas et al., 2021), leading to unrealistic and overly conservative behavior, while others treat robots as humans or add randomness for robustness (Chen C. et al., 2019; Stratton et al., 2024). However, these do not fully capture the diverse ways humans respond to robots, which depend on robot-specific factors like size, shape, and movement. More advanced crowd models that reflect these characteristics are needed for realistic social navigation simulation.

Advancing social navigation requires robust benchmarking methods that can accurately represent the planner’s performance. Key directions include adopting realistic crowd simulation, conducting real-world evaluations, and refining social metrics (Francis et al., 2023; Gao and Huang, 2022). Automated, objective real-world evaluation frameworks are increasingly important, as subjective user feedback is impractical to standardize. Future evaluations could use objective, non-verbal indicators, such as body language or facial expressions to better assess human comfort and social acceptance, ensuring planners are both effective and socially appropriate.

Social norms are informal rules guiding behavior in shared spaces, extending beyond collision avoidance and proxemics (Hall, 1963). For instance, smoothly avoiding social groups is addressed by some crowd prediction methods (Bisagno et al., 2018; Fernando et al., 2019), but is incorporated into only a few navigation algorithms (Bhaskara et al., 2023). Other norms, such as culturally specific conventions (Chen et al., 2017c), are context-sensitive and not universal, suggesting the value of learning social norms directly from large-scale crowd data rather than relying solely on handcrafted heuristics. Vision-language models (VLMs) open an additional pathway by enabling robots to ground these norms in natural language, reason about complex social contexts, and even communicate intentions to humans in interpretable ways. Effective social navigation will likely require a combination of data-driven norm learning and VLM-based reasoning, alongside intention communication that may be verbal (Dugas et al., 2020; Nishimura and Yonetani, 2020) or conveyed through non-verbal cues, as highlighted in autonomous vehicle research (Habibovic et al., 2018).

Social navigation is not a one-size-fits-all solution. Individuals and crowds vary in preferred comfort distance, speed, and interaction style. Future work should emphasize preference-aware navigation, where robots learn and adapt to individual users or cultural groups, potentially combining reinforcement learning with preference learning, feedback, or large language models that capture human expectations and feedback. Although current approaches consider human preferences during training (Choi et al., 2020), accommodating post-deployment feedback and achieving continuous learning remain open challenges.

A major challenge in learning-based planners is the difficulty of interpreting the reasoning behind their decisions, which is often referred to as explainability (Vouros, 2022). Integrating explainability improves user trust, allows better debugging, and clarifies the decision-making process. Several techniques exist, such as saliency maps, which visually indicate influential regions within image-based inputs (Huber et al., 2021), and approaches that provide verbal explanations for their decisions (Dugas et al., 2020). Integrating these explainability methods into learning-based social navigation can create more transparent, interpretable, and user-friendly systems.

Recent advances in vision-language navigation (VLNs) (An et al., 2022) highlight opportunities to enrich social navigation with multimodal reasoning capabilities and improve functional versatility. Beyond instruction following (Anderson et al., 2018a), VLNs can support a wide range of tasks such as visual question answering (Wu et al., 2024), describing social situations, or embodied dialog (Hahn et al., 2020). Social VLN could allow robots to interpret human intent, infer social norms from linguistic context, and communicate their own decisions in interpretable ways.