SocialCorridorWorld in Corridor-based AI Research

• SocialCorridorWorld is a multifaceted term describing various corridor-structured environments in computational research used to study prosocial behavior, navigation, and social dynamics.
• Different approaches employ minimal artificial-life rescue setups, graph-based robot navigation, and BDI-based indoor simulations to evaluate agent interactions under spatial constraints.
• Practical insights reveal that corridor topology amplifies trade-offs in progress, safety, and resource sharing, guiding experimental design in multi-agent and social simulation studies.

Searching arXiv for the cited papers and topic usage to ground the article in the current literature. SocialCorridorWorld is a polysemous label used in arXiv-indexed research to denote several distinct corridor-centered computational settings rather than a single canonical benchmark. In one usage, it names a minimal artificial-life rescue environment introduced in “Prosociality by Coupling, Not Mere Observation: Homeostatic Sharing in an Inspectable Recurrent Artificial Life Agent” (Sanyal, 12 Apr 2026), where a recurrent agent must decide whether to transport food to a passive partner under strictly self-directed homeostatic scoring. In another, it refers to the “Corridor” mini-game in SocialGym 2.0 (Sprague et al., 2023), a multi-agent social robot navigation benchmark in a narrow hallway. A third usage appears as a corridor-centric extension of IndoorWorld (Wu et al., 14 Jun 2025), emphasizing joint physical and social simulation for heterogeneous agents. Related corridor models also arise in pedestrian-flow simulation with dyadic group cohesion (Crociani et al., 2017), and the term has additionally been used for a multilayer virtual-world diffusion dataset (Jankowski et al., 2017). The common denominator is a corridor-structured domain in which spatial constraint, interaction, and coordination become experimentally salient, but the underlying formalism, task, and evaluation criteria differ substantially.

1. Terminological scope and research lineages

The most technically specific use of SocialCorridorWorld is the artificial-life environment in (Sanyal, 12 Apr 2026). There, the world is a one-dimensional corridor with explicit energetics, a passive partner, food transport, a static hazard, and a finite planning horizon. The environment was designed to test a narrow claim: helping appears when another’s need is routed into self-regulation, not when the agent merely observes a partner.

A second lineage comes from social robot navigation. In SocialGym 2.0, the relevant environment is the “Corridor” mini-game, instantiated in code via GraphNavScenario('envs/scenario/hallway') (Sprague et al., 2023). It is a long narrow rectangle with graph-structured traversal, partial observations, discrete high-level actions GO and [STOP](https://www.emergentmind.com/topics/self-taught-optimizer-stop), and a composite reward for progress, efficiency, and collision avoidance. Here the corridor is not a rescue domain but a shared-human-space navigation benchmark.

A third lineage is the corridor extension of IndoorWorld (Wu et al., 14 Jun 2025). In that formulation, SocialCorridorWorld denotes a graph-based environment with corridor segments, intersections, meeting spots, congestion, negotiations, and a Belief–Desire–Intention wrapper augmented by a Social Module. The emphasis is on physically grounded social simulation rather than minimal homeostatic control or classical navigation policy learning.

The term also intersects with crowd dynamics. Crociani et al. study a discrete toroidal corridor for uni-directional pedestrian flow, including dyads and a cohesion mechanism parameterized by $\kappa_c$ and $\delta$ (Crociani et al., 2017). Although the underlying paper is not framed around artificial agents with planning or recurrent control, it contributes an important corridor-based formalism for social grouping effects. By contrast, the Timik.pl dataset paper concerns five spreading campaigns over a multilayer virtual-world network (Jankowski et al., 2017); in that context, SocialCorridorWorld refers not to physical corridor locomotion but to a dataset/report label attached to diffusion and interaction structure.

This multiplicity of usage implies that SocialCorridorWorld should be treated as a family resemblance term. A plausible implication is that the label functions as a corridor-centered experimental motif across subfields rather than a standardized benchmark specification.

2. Artificial-life SocialCorridorWorld in homeostatic prosociality research

In (Sanyal, 12 Apr 2026), SocialCorridorWorld is formally specified as a 1d “corridor” of 19 cells indexed $0\ldots18$. The agent starts at cell 9, food spawns at cell 0, there is a static hazard at cell 9, and the partner is at cell 18. The possessor’s internal state includes true energy $E^{\mathrm{true}}_t \in [0,1]$, model-estimate of own energy $E^{\mathrm{model}}_t \in [0,1]$, a small recurrent hidden state $h_t$, and affective proxies valence $v_t$ and arousal $a_t$. The partner has true energy $E^{\mathrm{other,true}}_t \in [0,1]$, which is unobserved directly, and an estimate $\hat E^{\mathrm{other}}_t$ that is available in some social conditions.

The action space is $\delta$0. Left and Right move the agent one cell at cost $\delta$1; Get picks up food if co-located with it; Eat consumes carried food and raises own true energy by $\delta$2; Pass transfers carried food to the partner if co-located, after which the partner eats immediately and gains $\delta$3; and Stay imposes minimal cost. Hazard cost $\delta$4 is incurred if the agent enters or remains on the hazard cell. Episode horizon is $\delta$5 steps, planner rollout depth is $\delta$6, initial possessor energy is $\delta$7, initial partner energy is $\delta$8, and the carried-food buffer is empty at $\delta$9 (Sanyal, 12 Apr 2026).

The energy dynamics are explicitly homeostatic:

$0\ldots18$0

The paper further gives the full coupled homeostat:

$0\ldots18$1

Default parameters are setpoint $0\ldots18$2, coupling strength $0\ldots18$3, homeostat-to-prediction gain $0\ldots18$4, and prediction-error learning rate $0\ldots18$5 (Sanyal, 12 Apr 2026). Distress is implicit: when $0\ldots18$6, partner distress is $0\ldots18$7.

The agent optimizes only its own internal regulation. The self-scoring function over a candidate rollout of length $0\ldots18$8 is

$0\ldots18$9

with fixed weights $E^{\mathrm{true}}_t \in [0,1]$0, $E^{\mathrm{true}}_t \in [0,1]$1, $E^{\mathrm{true}}_t \in [0,1]$2, and $E^{\mathrm{true}}_t \in [0,1]$3 in the social tasks (Sanyal, 12 Apr 2026). No partner-welfare reward term is introduced. Coupling modifies only the predicted internal state via $E^{\mathrm{true}}_t \in [0,1]$4, so helping can emerge if it improves self-regulation.

3. Experimental conditions, lesions, and quantitative outcomes

The core comparison in (Sanyal, 12 Apr 2026) is among four matched conditions. In the “none” condition, no partner-state input is given. In the direct-state conditions, the agent receives $E^{\mathrm{true}}_t \in [0,1]$5. In “affective” and “full” conditions, that estimate is injected into the homeostat calculation via $E^{\mathrm{true}}_t \in [0,1]$6; in “cognitive” and “none,” $E^{\mathrm{true}}_t \in [0,1]$7, so partner input has no effect on $E^{\mathrm{true}}_t \in [0,1]$8.

Evaluation uses five metrics defined explicitly in the paper: help rate, rescue latency, partner recovery rate, mutual viability, and self-cost of help (Sanyal, 12 Apr 2026). Help rate is the fraction of episodes in which the possessor executes the full Get–Carry–Pass sequence; rescue latency is the timestep of the Pass action, or $E^{\mathrm{true}}_t \in [0,1]$9 if no help occurs; partner recovery rate is the fraction of episodes where $E^{\mathrm{model}}_t \in [0,1]$0; mutual viability is the average of $E^{\mathrm{model}}_t \in [0,1]$1 and $E^{\mathrm{model}}_t \in [0,1]$2 across episodes.

Under low load, with $E^{\mathrm{model}}_t \in [0,1]$3, 64 seeds, and deterministic trajectories, the dissociation is exact. The self-only and partner-observing conditions never help, whereas the affectively coupled conditions always do (Sanyal, 12 Apr 2026).

Condition	Help / recovery / latency	Mutual viability
social_none / social_cognitive_direct	Help rate = 0.0; Partner recovery = 0.0; Rescue latency = 18	0.15
social_affective_direct / social_full_direct	Help rate = 1.0; Partner recovery = 1.0; Rescue latency = 9	0.3286

The same runs also report approximate final energies. In social_none and social_cognitive_direct, final possessor $E^{\mathrm{model}}_t \in [0,1]$4 and partner $E^{\mathrm{model}}_t \in [0,1]$5. In social_affective_direct and social_full_direct, final possessor $E^{\mathrm{model}}_t \in [0,1]$6 and partner $E^{\mathrm{model}}_t \in [0,1]$7 (Sanyal, 12 Apr 2026). This establishes a self-cost of helping alongside improved joint survival. The paper’s abstract compresses the main result as follows: coupling flips help rate and partner recovery from 0 to 1, cuts rescue latency from 18 to 9 steps, and raises mutual viability from 0.15 to 0.33 (Sanyal, 12 Apr 2026).

Lesion experiments further constrain interpretation. Three conditions are tested: sham, coupling_off, and shuffle_partner. Sham preserves intact coupling; coupling_off forces $E^{\mathrm{model}}_t \in [0,1]$8 to 0 mid-trial; and shuffle_partner feeds a random partner-energy trace instead of true $E^{\mathrm{model}}_t \in [0,1]$9 (Sanyal, 12 Apr 2026). In both FoodShareToy and SocialCorridorWorld, sham lesions preserve helping, while coupling_off and shuffle_partner abolish it: help rate falls to 0, partner recovery to 0, and rescue latency returns to the horizon.

These manipulations address a common misconception: mere access to partner state is not sufficient in this architecture. The reported pattern distinguishes direct-state observation from affective incorporation into the homeostat.

4. Load dependence and feasibility boundary

SocialCorridorWorld in (Sanyal, 12 Apr 2026) includes three metabolic load regimes—low, medium, and high—implemented by scaling basal cost $h_t$0, move cost $h_t$1, hazard cost $h_t$2 upward and food gains $h_t$3 downward. The numeric table typically used in code is low: $h_t$4, medium: $h_t$5, and high: $h_t$6 (Sanyal, 12 Apr 2026).

A coupling sweep is performed over $h_t$7 under each load, with help rate and mutual viability as metrics. The principal result is a load-dependent feasibility boundary. Under low load, help rate jumps from 0 to 1 at $h_t$8, and mutual viability rises from approximately 0.15 at $h_t$9 to approximately 0.33 at $v_t$0, though self-cost steepens at highest $v_t$1 (Sanyal, 12 Apr 2026). Under medium and high load, no tested value yields any help: help rate remains 0, and mutual viability stays near 0.04 and 0.024, respectively.

The paper summarizes this schematically through a feasibility function,

$v_t$2

In the reported experiments, the boundary in the $v_t$3 plane is nonzero only in the low-load region for $v_t$4 (Sanyal, 12 Apr 2026). This suggests that coupling is necessary but not sufficient in the presence of harsher metabolic constraints. A plausible implication is that prosocial rescue in this minimal world is jointly determined by representational routing and resource feasibility.

5. SocialCorridorWorld as a navigation benchmark in SocialGym 2.0

In SocialGym 2.0, the relevant corridor environment is one of five social mini-games: Open, Doorway, Hallway, Intersection, and Roundabout (Sprague et al., 2023). The corridor geometry is a long narrow rectangle of width $v_t$5 and length $v_t$6, with two parallel walls. The navigation graph consists of two parallel chains of nodes, one on each wall. If there are $v_t$7 total agents, there are $v_t$8 nodes on each side, and every node on side A is connected via a straight-line edge to the corresponding node on side B.

Each agent $v_t$9 is assigned a global path that starts on one side and ends on the symmetric node on the opposite side. Spawn locations $a_t$0 and goal locations $a_t$1 are exactly the graph coordinates of those nodes. By default, each robot’s continuous state is

$a_t$2

where $a_t$3 (Sprague et al., 2023). Partial observation augments each agent’s own state with a localized representation of nearby agents, assembled in code by the Observer. The high-level action space is discrete, $a_t$4, and a local planner converts this plus the navigation graph into continuous commands $a_t$5.

The transition function is

$a_t$6

with $a_t$7 realized by a local trajectory sampler and kinodynamic emulator. Example unicycle kinematics are

$a_t$8

Ackermann- or differential-drive dynamics can be swapped in via configuration files (Sprague et al., 2023).

The reward in hallway benchmarks is a composite term:

$a_t$9

with default weights $E^{\mathrm{other,true}}_t \in [0,1]$0, $E^{\mathrm{other,true}}_t \in [0,1]$1, $E^{\mathrm{other,true}}_t \in [0,1]$2, $E^{\mathrm{other,true}}_t \in [0,1]$3, and $E^{\mathrm{other,true}}_t \in [0,1]$4 (Sprague et al., 2023). Collision occurs when circular footprints overlap, $E^{\mathrm{other,true}}_t \in [0,1]$5, and deadlock is detected if over 100 steps the sum of all agents’ displacements is less than $E^{\mathrm{other,true}}_t \in [0,1]$6 m.

Training uses PPO via Stable Baselines3 with policy = PPO("MlpPolicy", env, n_steps=4096, batch_size=2048, ent_coef=0.0, γ=0.99), over a total of 1.25 million environment steps and a curriculum that moves from 3 agents to 4 and then 5 (Sprague et al., 2023). Domain randomization varies the number of agents from 3–5 during training, while evaluation tests 3, 4, 5, 7, and 10 agents. Reported metrics are success rate, collision rate, average navigation time, average stop time, and maximum $E^{\mathrm{other,true}}_t \in [0,1]$7. In the hallway environment, the best collision rate, 0.04, is achieved by CADRL, while “Only Local” finishes fastest on average at 482 steps but has a collision rate of 2.8% (Sprague et al., 2023).

This version of SocialCorridorWorld is therefore structurally different from the homeostatic rescue world. Its central object is policy learning under partial observability and kinodynamic constraints, not coupling-mediated prosociality.

6. Extensions to heterogeneous social simulation and crowd-flow models

The IndoorWorld extension (Wu et al., 14 Jun 2025) pushes the corridor motif into a heterogeneous multi-agent environment that tightly integrates physical and social dynamics. World representation is a spatial graph $E^{\mathrm{other,true}}_t \in [0,1]$8 whose nodes include rooms, corridor intersections, and spots such as water-dispenser alcoves, and whose edges are corridor segments and doorways. Each edge $E^{\mathrm{other,true}}_t \in [0,1]$9 has length $\hat E^{\mathrm{other}}_t$0, width $\hat E^{\mathrm{other}}_t$1, capacity $\hat E^{\mathrm{other}}_t$2, and dynamic load $\hat E^{\mathrm{other}}_t$3. State is factored as $\hat E^{\mathrm{other}}_t$4, where $\hat E^{\mathrm{other}}_t$5 includes conversation sessions, an influence graph $\hat E^{\mathrm{other}}_t$6, and pending negotiations.

Agents are wrapped in a BDI architecture with perception, memory, planning, action, task prioritization, and a dedicated Social Module. The POMDP is defined as $\hat E^{\mathrm{other}}_t$7, with $\hat E^{\mathrm{other}}_t$8 and reward

$\hat E^{\mathrm{other}}_t$9

The social term includes

$\delta$00

while congestion cost is

$\delta$01

Dynamic edge weights for pathfinding are given by

$\delta$02

(Wu et al., 14 Jun 2025). The environment includes actions such as initiate_chat, yield_way, and propose_meet_spot, as well as token-based negotiation at narrow resources. Metrics include throughput, average traversal time, congestion index, clustering coefficient of the agent-agent graph at intersections, social cohesion score, and queue waiting times.

A separate but related line is the pedestrian-flow corridor in (Crociani et al., 2017). There the corridor has width $\delta$03 m, discretized into 8 cells of size $\delta$04 m, and a periodic measurement window of length $\delta$05 m. Time step is $\delta$06 s and motion uses a Moore neighbourhood. For isolated pedestrians, movement probability depends on a utility $\delta$07 built from goal attraction, obstacle repulsion, proxemics, direction inertia, and overlapping terms. For grouped pedestrians, an additional cohesion term $\delta$08 is introduced:

$\delta$09

The model calibrates $\delta$10 and $\delta$11, and reports that the presence of dyads reduces velocities and specific flow at medium-high densities (Crociani et al., 2017). Representative results show average speed reductions from 2% at density $\delta$12 ped/m² to 17% at density $\delta$13 ped/m², and flow reductions of similar magnitude. The maximum specific flow drops from approximately $\delta$14 ped·m$\delta$15·s$\delta$16 to approximately $\delta$17 ped·m$\delta$18·s$\delta$19 when 50% of the crowd are dyads (Crociani et al., 2017).

Taken together, these formulations show that corridor worlds support several distinct research programs: homeostatic helping, MARL-based navigation, heterogeneous socially situated planning, and socially modulated pedestrian flow. The shared spatial motif does not imply shared semantics or metrics.

7. Conceptual significance and recurrent misconceptions

Across these literatures, the corridor is valuable because it compresses spatial degrees of freedom and amplifies interaction. Narrow geometry forces trade-offs among progress, safety, congestion, resource transport, and social proximity. In (Sanyal, 12 Apr 2026), that compression isolates a mechanistic question: whether prosocial action can emerge without explicit partner reward. In (Sprague et al., 2023), it creates multi-agent interference patterns relevant to shared human spaces. In (Wu et al., 14 Jun 2025), it provides a substrate for congestion-sensitive BDI planning and negotiation. In (Crociani et al., 2017), it sharpens the measurable effects of social grouping on speed and flow.

Several misconceptions are not supported by the reported data. First, in the artificial-life setting, partner-state observation alone does not produce helping; the self-only and partner-observing conditions never help, whereas affectively coupled conditions always do under the low-load default (Sanyal, 12 Apr 2026). Second, coupling does not guarantee rescue in all regimes: under medium and high loads, no tested $\delta$20 rescues the partner within horizon (Sanyal, 12 Apr 2026). Third, faster traversal is not synonymous with socially better navigation in SocialGym 2.0; “Only Local” finishes fastest on average in the hallway environment but does so with a comparatively high collision rate of 2.8% (Sprague et al., 2023). Fourth, ignoring group cohesion in corridor flow can overestimate throughput, since dyads reduce both speed and specific flow at higher densities (Crociani et al., 2017).

The broader significance of SocialCorridorWorld lies in this experimental versatility. The term spans minimal inspectable agents, robot navigation simulators, graph-based social worlds, and crowd models, yet all instantiate a controlled setting in which corridor topology turns latent social structure into observable behavior. A plausible implication is that corridor worlds persist in research not because they are simple in a reductive sense, but because they are structurally constrained enough to make interaction mechanisms experimentally legible.