Social and Information-Aware Planning

• Socially and information-aware planners are decision-making systems that merge social context and informational state using probabilistic and learning methods to navigate complex environments.
• They combine techniques from POMDPs, epistemic planning, and inverse reinforcement learning to handle uncertainty, fuse sensor data, and respect social norms.
• Applications in robotics and autonomous vehicles demonstrate improved safety, social compliance, and coordinated multi-agent behavior in dynamic settings.

A socially- and information-aware planner is a decision-making system that explicitly reasons about both the social context (beliefs, awareness, intentions of other agents) and its own informational state (uncertainty, observability, and knowledge fusion) when acting in environments shared with humans or other intelligent agents. These planners are core to modern robotics, autonomous vehicles, and human–robot interaction scenarios, enabling robots to perform not only safe but also socially considerate and contextually interpretable actions. Their formalizations synthesize methodologies from POMDPs, epistemic and multi-agent planning, inverse reinforcement learning, deep learning, and model-space search.

1. Formal Models and Problem Definitions

The core principle in socially- and information-aware planning is the extension of classic planning state spaces to encode both physical world features and latent social/information variables.

• POMDP Formulations: In social navigation (e.g., mobile robots among people), the state encompasses physical robot and pedestrian positions, but also latent variables like human awareness, each modeled as random variables (e.g., $p_{\text{awareness}}^i\in\{+1, -1\}$). The planner reasons over this hybrid and partially observable state, maintaining beliefs via particle filtering and Bayesian fusion (Kim et al., 2018).
• Epistemic Planning: In multi-agent epistemic frameworks, the formal problem captures world state $w$ and nested belief states $B_1,\dots,B_n$ (for agents 1 through $n$), enabling the agent to reason about other agents' knowledge, misperceptions, and social norms (Sonenberg et al., 2016).
• Multi-Model Human-Aware Task Planning: In task planning, both the agent's task model $M^R$ and an explicit model of the observer’s (human’s) mental model $M^R_h$ are maintained. Differences $\Delta(M^R, M^R_h)$ between them are measured and traded off during plan generation (Chakraborti et al., 2017).

2. Social Perception and Information Fusion

Socially-aware planners integrate advanced perception pipelines and belief-maintenance mechanisms:

• Sensor Fusion for Social Variables: Sensor streams (e.g., laser, vision, YOLO, gaze detectors) are fused to detect people, track positions, and infer awareness cues. Awareness is encoded based on gaze (face-oriented vs. not), with prediction and filtering (e.g., Kalman filters per trajectory) (Kim et al., 2018).
• Distributed Sensor Networks: In autonomous driving, all agents (other vehicles and pedestrians) are treated as distributed sensors, where their individual and group behaviors inform the robot's beliefs about the world and social dynamics. Bayesian filtering is performed over both physical variables (e.g., occlusions) and latent social parameters (local driving styles, courtesy norms) (Sun et al., 2019).
• Cooperative Infrastructure: Systems may leverage infrastructure sensor nodes, fusing multi-camera and sparse LiDAR data for precise 3D human pose estimation. Joint pose uncertainties are propagated from camera calibration and projection, with global belief fusion resolving asynchronous reports and occlusions (Ning et al., 8 Apr 2025).

3. Planning Algorithms and Social Objective Functions

The core of these planners is in policy construction under social/informational constraints.

• Socially-Aware POMDP Planning: Planners solve finite-horizon POMDPs (often with DESPOT or similar real-time solvers), where the reward function includes awareness-sensitive collision penalties:

$$R_{\text{col}}(s)=\sum_{i=1}^N \begin{cases} -C_{\text{col}}, & \|R_{pos}-p_{pos}^i\|\le \rho_i\ 0, & \text{otherwise} \end{cases}$$

with $\rho_i$ (personal space) strictly dependent on $p_{\text{awareness}}^i$ (Kim et al., 2018).

• Inverse RL and Social Cost Learning: Cost functions for model predictive control are learned via inverse reinforcement learning on human data, capturing not only efficiency and safety but social compatibility and local norms (e.g., speed, yielding) (Sun et al., 2019).
• Explicit Social Norms via Potential Fields: Some planners (e.g., SAP-CoPE) embed constraints to respect areas of discomfort around human bodies (learned ellipsoidal personal spaces derived from pose), and explicitly penalize traversals through these fields (Ning et al., 8 Apr 2025).
• Topology-Aware Path Planning: Homology classes of paths are used to represent high-level social navigation choices (e.g., passing on left/right), and a deep neural network trained on human data selects the path class most likely to match social conventions (Martinez-Baselga et al., 2024).
• K-Step Social Intrusion Prediction: Multi-robot planners discourage myopic or aggressive motion by predicting future proximity to human comfort zones (using pedestrian trajectory predictors) and incorporating this as a lookahead reward (He et al., 2022).

4. Multi-Agent and Cooperative Planning

Socially-aware planning is often fundamentally multi-agent, requiring sophisticated representations and coordination strategies:

• Decentralized POMDP and Multi-Agent RL: In scenarios with multiple robots, the planning problem is formalized as a Dec-POMDP, solved using CTDE-style off-policy multi-agent RL frameworks (e.g., MSA³C), allowing robots to coordinate trajectories under local, partial observability (He et al., 2022).
• Temporal-Spatial Graph Encoding: Robots encode observed humans (and other robots) in a temporal-spatial graph structure, using sequence models and attention mechanisms to focus prediction and policy updates on the most socially relevant agents (He et al., 2022).
• Global Attention in Value Estimation: Multi-head global attention modules in critic networks aggregate inter-robot and agent-pedestrian relationships to better inform decentralized policy updates, which is critical when robot density or pedestrian flows are high (He et al., 2022).

5. Human-Aware Task Planning: Explicability and Explanation

Information-awareness extends beyond local motion: planners engage in meta-reasoning about observer models, explicability, and post-hoc explanation.

• Model-Space Search: The MEGA algorithm searches in the space of possible human (observer) models, seeking task plans that are both explicable (requiring no or minimal explanation in the human model) and, when necessary, optimizing the trade-off between explanation length and robot-optimality via a joint objective:

$$\min_{\pi, E} \left[ |E| + \alpha |C(\pi, M^R) - C_{M^R}^*| \right]$$

(Chakraborti et al., 2017).

• Human Factors and Adaptivity: Empirical human-in-the-loop studies demonstrate substantial variance in users’ willingness to demand explanations, suggesting that online estimation of user-specific trade-off parameters ($\alpha$) is required for optimal interaction (Chakraborti et al., 2017).
• Epistemic Planning for Social Interaction: Social epistemic planners manipulate not only the task world but also the states of belief of self and others, performing information-sharing actions (e.g., signal, explain, announce), and planning not simply for physical outcomes but for desired social/informational states (e.g., mutual understanding) (Sonenberg et al., 2016).

6. Empirical Evaluation and Metrics

A range of quantitative and qualitative metrics are used to assess performance:

• Physical Safety and Efficiency: Collision rates, success (goal-reaching), travel time, and average path length are standard. Social planners demonstrably outperform baseline reactive approaches, especially in dense and ambiguous environments (Kim et al., 2018, He et al., 2022, Martinez-Baselga et al., 2024).
• Social Compliance and Comfort: Minimum distance to pedestrians, comfort intrusion rates, and explicit discomfort penalty rates measure adherence to learned or specified social zones (Kim et al., 2018, Ning et al., 8 Apr 2025).
• Social Intelligence Success: "Social intelligence" is assessed by the planner’s ability to match human navigation modes (homology classes), yielding, passing, and acting with social preference distribution similar to real crowds (Martinez-Baselga et al., 2024).
• Task-Level Interpretability: For human-aware planning, key metrics include explanation length, plan explicability gap, user click-through on explanations, and human-rated trust and comfort (Chakraborti et al., 2017).

7. Limitations and Future Directions

Despite significant progress, several open challenges remain:

• Scalability: Exact belief maintenance (in epistemic or POMDP spaces) becomes intractable with many agents or deep theory-of-mind, motivating research into approximation, stereotype reasoning, and learning-based heuristics (Sonenberg et al., 2016, He et al., 2022).
• Perception Integration: Most state spaces assume perfect (or at least robust) perception; in practice, sensor failures, occlusions, and imperfect pose estimation degrade model accuracy. Joint perception–planning optimization and real-world robustness are active research areas (Ning et al., 8 Apr 2025, He et al., 2022).
• Norm Generalization and Learning: Social norm libraries are typically hand-coded or learned from limited demonstration data. Extending to diverse/generative social settings and learning norms adaptively remains an open direction (Sonenberg et al., 2016, Sun et al., 2019).
• Real-Time and Real-World Deployment: Maintaining high update rates in dense human environments, with real-time constraints and degraded sensing, is not universally solved. Most sophisticated planners report online feasibility for moderate robot/human counts but note computational constraints (Kim et al., 2018, He et al., 2022, Ning et al., 8 Apr 2025).
• Human Factors and Cognitive Costs: Explicit modeling of human observer costs (processing explanations, reassessment efforts) and combined cognitive–social reasoning in planning objectives are largely unexplored outside laboratory studies (Chakraborti et al., 2017).

Socially- and information-aware planning thus unifies rich environment representation, probabilistic inference, multi-agent coordination, and human-adaptive reasoning, demonstrated across domains from service robots and autonomous vehicles to HRI and task-level collaboration (Kim et al., 2018, Sonenberg et al., 2016, Chakraborti et al., 2017, Sun et al., 2019, He et al., 2022, Ning et al., 8 Apr 2025, Martinez-Baselga et al., 2024).