S‑NAMO: Socially-Aware Navigation

• S‑NAMO is a navigation paradigm that augments classical NAMO by incorporating social constraints and semantic rules for obstacle movability and placement.
• It uses semantic white-listing and taboo zones to filter obstacles, ensuring that object displacement minimally disturbs human activity.
• The framework leverages multi-objective planning and robust trajectory prediction to achieve safe, socially compliant navigation in complex human environments.

Searching arXiv for the cited S‑NAMO and related socially-aware navigation papers.

Socially-Aware Navigation Among Movable Obstacles (S‑NAMO) is an extension of the classical Navigation Among Movable Obstacles (NAMO) problem that incorporates social constraints and human acceptability into navigation and manipulation decisions. In the formulation introduced by “Towards S‑NAMO: Socially-aware Navigation Among Movable Obstacles” (Renault et al., 2019), the objective is to allow an autonomous agent to go from an initial pose to a goal pose, forbidding collision with obstacles but not their displacement, while minimizing both robot’s displacement cost and disturbance to humans. Within the provided literature, S‑NAMO appears both as a concrete algorithmic proposal centered on social movability evaluation and social placement of objects (Renault et al., 2019), and as a broader design problem connected to socially-aware trajectory prediction (Saadatnejad et al., 2021), unified socially-aware local planning (Banisetty et al., 2019, Banisetty et al., 2021), negotiation-based navigation (d'Addato et al., 2024), interaction-triggered navigation (Cruz-Maya et al., 2019), language-mediated navigation (Wen et al., 2024), contact-aware planning among movable objects (Wang et al., 5 Feb 2025), and formally assured socially-aware decision-making (Lestingi et al., 30 Jun 2026).

1. Classical NAMO and the definition of S‑NAMO

Classical NAMO, formally introduced by Stilman et al. (2005), assumes a robot operating in a workspace with static and movable obstacles; the robot may move some obstacles, by pushing or grasping, to clear a path to a goal (Renault et al., 2019). In the representation described for this setting, the state includes robot pose and obstacle positions,

$s = (q_r, O),\quad O = \{o_1,\dots,o_m\},$

and plans are evaluated with navigation-oriented costs such as distance, time, energy, number of moved obstacles, and probability of success (Renault et al., 2019). The literature surveyed in that work emphasizes that many NAMO variants are NP-hard or PSPACE-hard, so most planners pursue completeness guarantees or locally optimal variants rather than global optimality (Renault et al., 2019).

S‑NAMO is defined by adding social acceptability to that formulation. The paper states: “We want to create Social NAMO algorithms: ones that allow an autonomous agent to go from an initial pose to a goal pose, forbidding collision with obstacles but not their displacement, minimizing both robot's displacement cost (distance, time or energy) and disturbance to humans” (Renault et al., 2019). A natural formalization given in the provided material is

$$\min_{\pi} J(\pi) = \sum_{t} \Big( C_{\text{nav}}(s_t,a_t) + \lambda\, C_{\text{social}}(s_t,a_t) \Big),$$

subject to collision constraints, physical feasibility of manipulation, and additional social constraints on which obstacles can be moved and where they may be placed (Renault et al., 2019).

The initial S‑NAMO proposal identifies three conceptual objectives. The first is Social Movability Evaluation, which asks whether an object is socially acceptable to move. The second is Social Placement Choice, which asks whether the chosen final placement of the object is socially acceptable. The third is Social Action Planning, which asks whether the robot’s physical actions themselves are socially acceptable in the presence of humans; the paper explicitly leaves this third component for future work and focuses on the first two under delayed human-object interaction, meaning humans are not physically present while the robot operates but will later use the environment (Renault et al., 2019).

This decomposition is central because it shifts NAMO away from a purely geometric treatment of obstacles. In the S‑NAMO view, a movable obstacle is not just an object with a kinematic affordance; it is also a socially situated entity whose displacement may interfere with ownership, circulation, visibility, comfort, or downstream human use. A plausible implication is that S‑NAMO is best understood not as a single planner, but as a layered decision problem combining motion planning, semantic interpretation, and normative filtering.

2. Social movability evaluation and socially acceptable placement

The first concrete S‑NAMO extension is a social movability filter. In the baseline Wu–Levihn NAMO planner on which the paper builds, any obstacle is initially considered movable until a failed manipulation attempt blacklists it (Renault et al., 2019). S‑NAMO replaces this with a white-listing mechanism driven by semantic perception. The robot is assumed to have a geometric sensor with a wide field of view and a semantic sensor with a narrower field of view; the geometric sensor detects and segments obstacles, while the semantic sensor recognizes and classifies them as movable or unmovable when they enter its field of view (Renault et al., 2019).

This yields three obstacle states: known movable, known unmovable, and unknown (Renault et al., 2019). Unknown obstacles are treated as unmovable unless the planner finds a route through a pose where the object enters the semantic field of view and can be classified. The key condition added to the algorithm is: $$\texttt{if } \texttt{is-movable}(o) \text{ or } \big( \texttt{is-unknown}(o) \text{ and } c_0 \neq \emptyset \text{ and } c_1 \neq \emptyset \big),$$ which means that a manipulation plan is only evaluated if the obstacle is already known movable, or if it is unknown but the planner has found a feasible observation-and-approach decomposition (Renault et al., 2019). In effect, social movability modifies obstacle selection before manipulation planning begins.

The second extension is social placement. The world model is augmented with a semantic map layer containing taboo polygons,

$P = \{P_1, \dots, P_k\},$

which designate socially forbidden placement regions such as doorways, circulation zones, or a valuable carpet (Renault et al., 2019). The admissible placements for an object $o$ are then expressed as

$$\mathcal{P}_{\text{allowed}}(o) = \{ p \in \mathcal{C}_r \mid \text{footprint}(o,p) \cap P = \emptyset \}.$$

In implementation terms, placement is accepted only if opening creation holds and the predicate not-in-taboo(w,o) is satisfied (Renault et al., 2019).

The social reasoning in this first S‑NAMO formulation is intentionally simple and binary. Social movability is effectively a yes/no classification, and placement feasibility is effectively a taboo/non-taboo test (Renault et al., 2019). The paper explicitly notes that there is no scalar social movability score and no richer placement utility. This is important because it clarifies what S‑NAMO initially contributes: not a complete normative model, but a minimal operational mechanism for excluding socially problematic object moves and placements from the planner’s search space.

A concise summary of the two implemented S‑NAMO extensions is useful.

Component	Mechanism	Role in planning
Social Movability Evaluation	Semantic white-listing of movable objects	Filters candidate obstacles
Social Placement Choice	Taboo polygons in a semantic layer	Filters candidate placements

The simulator scenario described in the paper illustrates the effect of these constraints. In a two-room environment with two boxes blocking doorways, standard NAMO may move one box just enough to pass through the first doorway, leaving it in a position that blocks the second doorway. S‑NAMO, by contrast, pushes boxes outside the doorway regions and outside taboo zones, enabling both goals to be reached while keeping doorway areas clear (Renault et al., 2019).

3. Semantic layers, social rules, and social cost structure

S‑NAMO depends on semantics because social acceptability is rarely recoverable from geometry alone. The initial implementation uses a 2D polygonal map, grid-based A* for path planning, polygonal movable obstacles, and a semantic map layer for taboo zones (Renault et al., 2019). In abstract form, the state can be regarded as

$s = (x_r, O, S),$

where $x_r$ is robot pose, $O$ is the set of objects, and $S$ is the semantic/social layer (Renault et al., 2019). In that layer, taboo zones determine where objects must not be placed, while object recognition determines which objects are socially movable.

Later socially-aware navigation work broadens this semantic perspective. Unified socially-aware planners based on PaCcET treat proxemics, queue joining, group interaction, and activity-space avoidance as separate objectives rather than as a single weighted sum (Banisetty et al., 2019, Banisetty et al., 2021). The survey literature characterizes socially aware navigation as requiring at least human detection and special treatment, plus comfort and non-disturbance (Singamaneni et al., 2023). In this language, S‑NAMO adds object semantics and object-placement semantics to the same stack of social constraints.

Several concrete cost structures from related work are directly relevant. A typical navigation objective vector may combine goal progress, obstacle avoidance, and social terms such as interpersonal distance, group distance, social goal distance, or activity-space avoidance (Banisetty et al., 2019). The social distance terms described in that line of work include

$$\min_{\pi} J(\pi) = \sum_{t} \Big( C_{\text{nav}}(s_t,a_t) + \lambda\, C_{\text{social}}(s_t,a_t) \Big),$$0

where $$\min_{\pi} J(\pi) = \sum_{t} \Big( C_{\text{nav}}(s_t,a_t) + \lambda\, C_{\text{social}}(s_t,a_t) \Big),$$1 is robot–person distance, $$\min_{\pi} J(\pi) = \sum_{t} \Big( C_{\text{nav}}(s_t,a_t) + \lambda\, C_{\text{social}}(s_t,a_t) \Big),$$2 is distance to a group centroid or boundary, and $$\min_{\pi} J(\pi) = \sum_{t} \Big( C_{\text{nav}}(s_t,a_t) + \lambda\, C_{\text{social}}(s_t,a_t) \Big),$$3 is distance to a context-specific social goal (Banisetty et al., 2019). For S‑NAMO, the natural extension is to treat moved objects as socially relevant entities as well, so that object–human distance, object final placement, and social disruption enter the objective vector alongside ordinary navigation terms. This suggestion is explicit in the PaCcET-based SAN literature, which recommends adding manipulation effort, ownership penalties, and human disruption costs when extending socially-aware navigation to NAMO (Banisetty et al., 2019).

The 2023 survey also identifies environment semantics, object interactions, human intentions, and social norms as separate facets of socially aware navigation (Singamaneni et al., 2023). That framing aligns closely with S‑NAMO’s needs. A plausible implication is that fully developed S‑NAMO systems require at least four semantic distinctions: whether an object is physically movable, whether it is socially movable, where it may be placed, and how its movement affects nearby human activities.

4. Planning architectures relevant to S‑NAMO

The earliest S‑NAMO paper is largely conceptual and attaches its extensions to an existing NAMO planner (Renault et al., 2019). Subsequent socially-aware navigation research offers several architectures that can serve as local or high-level modules inside S‑NAMO.

One prominent line is the unified socially-aware navigation architecture built around context classification and PaCcET multi-objective local planning (Banisetty et al., 2019, Banisetty et al., 2021). In that architecture, a context classifier identifies hallway, art gallery, queue, group, or similar scenarios; the planner then activates a context-specific set of social objectives and uses PaCcET to choose a locally optimal trajectory (Banisetty et al., 2021). This is relevant to S‑NAMO because movable obstacles introduce additional contexts: narrow passages with chairs, waiting areas with carts, shared workspaces with personal belongings, or emergency corridors with temporarily movable debris. The same architecture can therefore be extended by enriching the context classifier and by adding NAMO-specific objectives such as number of obstacles moved, displacement magnitude, or taboo-zone avoidance (Banisetty et al., 2021).

A second line emphasizes negotiation before geometry. “Socially-Aware Opinion-Based Navigation with Oval Limit Cycles” couples nonlinear opinion dynamics with vortex fields and oval limit cycles (d'Addato et al., 2024). The opinion variable $$\min_{\pi} J(\pi) = \sum_{t} \Big( C_{\text{nav}}(s_t,a_t) + \lambda\, C_{\text{social}}(s_t,a_t) \Big),$$4 encodes left/right passing preference, attention $$\min_{\pi} J(\pi) = \sum_{t} \Big( C_{\text{nav}}(s_t,a_t) + \lambda\, C_{\text{social}}(s_t,a_t) \Big),$$5 rises as the encounter becomes critical, and a pitchfork bifurcation forces a side choice when ambiguous encounters must be resolved (d'Addato et al., 2024). The paper explicitly relates this to S‑NAMO by suggesting that the same opinion dynamics can be extended from pass-left/pass-right to move-obstacle/bypass/wait decisions, with non-verbal cues translated into an interaction opinion and consensus used to avoid deadlocks (d'Addato et al., 2024). That is not an implemented S‑NAMO system in the source paper, but it is an explicit conceptual bridge in the provided material.

A third line inserts human–robot interaction into the navigation loop. In “Enabling Socially Competent navigation through incorporating HRI,” the system monitors local planner failure, clusters nearby obstacles, detects whether the blockage is human, compares alternative path length against the original path, and, if the detour is socially undesirable, interrupts navigation to ask for permission to pass (Cruz-Maya et al., 2019). This is a negotiation policy rather than a manipulation planner, but it is directly NAMO-like in the sense that blocked space is no longer treated as a fixed geometric constraint; the robot can transform obstruction into interaction.

More recent work pushes this interaction further. HSAC-LLM uses Hybrid Soft Actor-Critic with a joint continuous–discrete action space, where continuous actions are velocity commands and discrete actions are interaction codes such as none, stop, margin-right, and margin-left (Wen et al., 2024). The language stack both encodes incoming human utterances and generates outgoing requests, enabling bidirectional natural-language coordination during socially-aware navigation (Wen et al., 2024). The paper explicitly notes that these discrete interaction actions can be viewed as commands that reconfigure human positions, which is structurally close to a social version of NAMO (Wen et al., 2024).

These architectures differ in mechanism—multi-objective optimization, dynamical systems, HRI triggering, RL plus language—but they share a common S‑NAMO lesson: the planner must reason about more than collision-free geometry. It must reason about context, social negotiation, and the acceptability of changing the environment.

5. Prediction, robustness, and human-group structure

Trajectory prediction is a core dependency in many socially-aware planners, and the robustness of that prediction is directly relevant to S‑NAMO. “Are socially-aware trajectory prediction models really socially-aware?” studies six representative predictors—S‑LSTM, Social-Attention, Social-GAN, Directional-Pooling, Social-STGCNN, and PECNet—and evaluates social understanding through collision avoidance (Saadatnejad et al., 2021). The paper defines a collision between candidate agent 1 and neighbor $$\min_{\pi} J(\pi) = \sum_{t} \Big( C_{\text{nav}}(s_t,a_t) + \lambda\, C_{\text{social}}(s_t,a_t) \Big),$$6 at prediction time $$\min_{\pi} J(\pi) = \sum_{t} \Big( C_{\text{nav}}(s_t,a_t) + \lambda\, C_{\text{social}}(s_t,a_t) \Big),$$7 as

$$\min_{\pi} J(\pi) = \sum_{t} \Big( C_{\text{nav}}(s_t,a_t) + \lambda\, C_{\text{social}}(s_t,a_t) \Big),$$8

with $$\min_{\pi} J(\pi) = \sum_{t} \Big( C_{\text{nav}}(s_t,a_t) + \lambda\, C_{\text{social}}(s_t,a_t) \Big),$$9 m, and reports Collision Rate (CR) as the percentage of scenarios with at least one such collision (Saadatnejad et al., 2021).

The key empirical finding is that even without attack, these “social” predictors often output colliding trajectories: S‑LSTM 7.8%, S‑Att 9.4%, D‑Pool 7.3%, S‑GAN 13.9%, S‑STGCNN 16.3%, and PECNet 15.0% (Saadatnejad et al., 2021). Under soft-attention S‑ATTack with realistic perturbations bounded by $$\texttt{if } \texttt{is-movable}(o) \text{ or } \big( \texttt{is-unknown}(o) \text{ and } c_0 \neq \emptyset \text{ and } c_1 \neq \emptyset \big),$$0 m, CR rises sharply—to 89.8% for S‑LSTM, 86.4% for S‑Att, 85.0% for S‑GAN, 88.0% for D‑Pool, 59.1% for S‑STGCNN, and 64.9% for PECNet—while average perturbation magnitudes remain small, roughly 0.03–0.11 m (Saadatnejad et al., 2021). The paper argues that this shows limited collision reasoning and high vulnerability of trajectory predictors under realistic localization error. For S‑NAMO, that matters because planning around humans or deciding when an obstacle will remain unused depends on multi-agent forecasts; brittle predictors can induce socially inconsistent or unsafe manipulations.

The same paper also demonstrates that adversarially generated examples can be used for robustification. On D‑Pool, adversarial training reduces original CR from 7.3% to 6.5% and reduces CR under attack from 37.3% to 14.7% (Saadatnejad et al., 2021). This suggests that S‑NAMO systems using learned prediction modules should treat social robustness as a first-class requirement, not merely optimize ADE/FDE.

Human-group structure is another missing piece in many planners. SANGO addresses this by clustering dynamic obstacles using DBSCAN and distinguishing core group members $$\texttt{if } \texttt{is-movable}(o) \text{ or } \big( \texttt{is-unknown}(o) \text{ and } c_0 \neq \emptyset \text{ and } c_1 \neq \emptyset \big),$$1 from boundary members $$\texttt{if } \texttt{is-movable}(o) \text{ or } \big( \texttt{is-unknown}(o) \text{ and } c_0 \neq \emptyset \text{ and } c_1 \neq \emptyset \big),$$2 (Malladi et al., 2024). The reward function imposes a stronger penalty for collisions with group core points than with individual dynamic obstacles, and uses inverse-distance or exponential penalties within personal-space and group-boundary thresholds (Malladi et al., 2024). In the reported experiments, grouping reduces discomfort by up to 83.5%, collision rates by up to 29.4%, and improves success in crowded scenarios (Malladi et al., 2024). Although SANGO does not perform NAMO, it provides a directly relevant lesson: movable-object planning in crowds must respect groups as social entities, not just as multiple independent pedestrians.

6. Toward advanced S‑NAMO: contact, abstractions, and formal guarantees

Beyond the initial S‑NAMO proposal, several later works indicate how the field can evolve technically.

Contact-Aware Motion Planning Among Movable Objects (CAMP) formulates NAMO and RAMO as a joint spatial-temporal trajectory optimization problem over robot and object states, with contact encoded through complementarity constraints and solved by augmented Lagrangian methods (Wang et al., 5 Feb 2025). The full optimization includes boundary conditions, collision avoidance, velocity and acceleration bounds, system dynamics, and complementarity constraints, with an objective that combines smoothness, total duration, and a task term (Wang et al., 5 Feb 2025). In benchmark NAMO tasks, CAMP achieves a total average success rate of 95% versus 54.3% for a collision-free baseline, albeit with higher computation time (Wang et al., 5 Feb 2025). The provided material explicitly suggests that CAMP provides natural hooks for S‑NAMO through additional social costs such as human proximity, socially sensitive object displacement, and path-blocking penalties (Wang et al., 5 Feb 2025). This is interpretive in the source, but it is a clear design direction grounded in the optimization structure.

A different abstraction route comes from context-specific abstract MDPs (CAMPs). That work treats planning domains, including NAMO, as factored MDPs and learns self-imposed constraints that induce context-specific independences, allowing irrelevant variables and actions to be dropped from the planning problem (Chitnis et al., 2020). The objective balances reward and computational cost,

$$\texttt{if } \texttt{is-movable}(o) \text{ or } \big( \texttt{is-unknown}(o) \text{ and } c_0 \neq \emptyset \text{ and } c_1 \neq \emptyset \big),$$3

and planning in learned CAMPs outperforms several baselines, including Stilman’s NAMO-specific algorithm, in the reported NAMO domain (Chitnis et al., 2020). A plausible implication is that S‑NAMO could use socially defined contexts—such as “do not enter interaction zone” or “avoid rooms with occupied furniture”—to induce compact socially constrained abstractions before motion planning begins.

Formal assurance is addressed by FormIDEAble, which models socially-aware cooperation as a Priced Timed Markov Decision Process and synthesizes strategies via a cost-bounded reachability problem (Lestingi et al., 30 Jun 2026). The decision problem is written as

$$\texttt{if } \texttt{is-movable}(o) \text{ or } \big( \texttt{is-unknown}(o) \text{ and } c_0 \neq \emptyset \text{ and } c_1 \neq \emptyset \big),$$4

In the emergency-evacuation example, the framework encodes uncertainty in human cooperation, prioritization of vulnerable individuals, and resource bounds, then synthesizes strategies with safety guarantees (Lestingi et al., 30 Jun 2026). The paper does not address geometric NAMO directly, but the mapping to S‑NAMO is explicit in the provided material: high-level choices such as asking a human to move, choosing an alternate corridor, or manipulating an object can be lifted into a timed probabilistic decision model with formal safety and social constraints (Lestingi et al., 30 Jun 2026).

These directions suggest that mature S‑NAMO may combine at least three levels: a semantic-social layer deciding what is permissible, a geometric-dynamic layer deciding how to move in continuous space, and a verification or robust-planning layer certifying that social and safety constraints hold under uncertainty.

7. Limitations, misconceptions, and research directions

A common misconception is that any NAMO system operating in human environments becomes socially aware once it avoids collisions. The sources do not support that view. The original S‑NAMO paper explicitly argues that existing NAMO literature had not directly addressed social constraints and treated movable obstacles largely as geometric entities with a movability flag (Renault et al., 2019). Socially aware navigation work similarly distinguishes mere collision avoidance from richer notions of comfort, social conflict resolution, intent exhibition, and context-dependent behavior (Singamaneni et al., 2023).

A second misconception is that social acceptability can be reduced to one rule, such as minimum interpersonal distance. The provided works collectively show a broader landscape: social placement, queue semantics, O-formation geometry, activity-space avoidance, passing-side conventions, negotiation, group cohesion, prediction robustness, and task- or role-dependent priorities all matter (Renault et al., 2019, Banisetty et al., 2019, Banisetty et al., 2021, d'Addato et al., 2024, Malladi et al., 2024, Lestingi et al., 30 Jun 2026).

The limitations of current S‑NAMO proposals are also explicit. The 2019 S‑NAMO paper uses a simple binary model of movability and taboo zones, assumes delayed human-object interaction, assumes perfect perception, and uses simple pushing of convex polygons in 2D (Renault et al., 2019). Pepper experiments in that work show that pushing light bins is feasible, but heavier or caster-equipped objects such as chairs can drift unpredictably, motivating more realistic manipulation models and uncertainty handling (Renault et al., 2019). Related socially-aware navigation systems often remain reactive, are evaluated in relatively simple single-human scenarios, or lack explicit treatment of object ownership, group semantics, or real-world crowd structure (Cruz-Maya et al., 2019, d'Addato et al., 2024).

The future directions present in the provided literature converge on a coherent S‑NAMO agenda. They include richer semantic and social layers, direct and indirect human interaction, robust prediction under perturbation, context-sensitive objective selection, multi-agent negotiation, language-mediated coordination, contact-aware optimization, and formal safety guarantees (Renault et al., 2019, Saadatnejad et al., 2021, Banisetty et al., 2021, Wen et al., 2024, Wang et al., 5 Feb 2025, Lestingi et al., 30 Jun 2026). A plausible implication is that S‑NAMO will ultimately require integration rather than replacement: semantic perception, social prediction, planning among movable objects, and socially grounded decision-making must be composed into a single system.

In that sense, S‑NAMO designates not just a variant of rearrangement planning, but a broader research program. It starts from the observation that path clearance in human spaces is never purely geometric, because what may be moved, where it may be moved, and whether it should be moved at all depend on human norms, activities, expectations, and safety. The literature in the provided corpus establishes the first operational definitions and mechanisms for that claim (Renault et al., 2019), and the surrounding socially-aware navigation literature supplies the prediction, optimization, interaction, and verification tools needed to make it technically credible.