PaperlayerView exact revision on arXiv

Source-linked AI summary

Networked Humanoid Intelligence: Communication, Coordination, Collaboration, and Social Interaction

Loc X. Nguyen, Huy Q. Le, Ji Su Yoon, Nguyen H. Tran, Eui-Nam Huh, Choong Seon Hong

arXiv:2610.34740v1cs.ROcs.ETcs.NI

TL;DR

Humanoid robotics research has largely treated robots as individual systems, despite the growing importance of information exchange across connected fleets. This survey organizes networked humanoid intelligence into four functions and reviews evidence across platforms and embodiment levels. It finds uneven evidence for humanoid-specific collaboration and social interaction, alongside open needs for compact body-centric messages and timely human oversight.

  • Problem

    Research has largely studied humanoids as individual systems, leaving limited evidence on how networked functions operate under humanoid constraints.

    Source: §1 · p. 2, §1.1 · p. 2

  • Method

    The survey organizes networked humanoid intelligence into communication, coordination, collaboration, and social interaction, comparing evidence across functions and embodiment levels.

    Source: §8 · p. 29

  • Results

    Reasoning and capability composition are tested only with simulated agents or robotic arms, experience sharing with wheeled robots, and multi-humanoid social interaction has limited evidence.

    Source: §5.4 · p. 22, §6.4 · p. 25

  • Takeaways & Limitations

    Compact body-centric messages could help robots assess whether teammates’ plans remain physically compatible without exchanging complete trajectories.

    Source: §7.4 · p. 28

    Humanoid fleets still lack compact representations for body-centric information and how long that information remains valid.

    Source: §3.5 · p. 13

Abstract

from arXiv · show

Humanoid robots are moving from laboratory platforms toward connected fleets that share workspaces, infrastructure, tasks, and knowledge with other robots and humans. As deployments grow, the operation and improvement of each humanoid depend on information exchanged across the network. However, existing research has largely studied humanoids as individual systems. This survey examines networked humanoid intelligence through four functions of information exchange: communication, coordination, collaboration, and social intelligence with human interaction. We first establish the foundations of networked humanoid systems, define the question addressed by each function, and connect them through a common factory scenario. We review the literature along six dimensions: interaction architecture, exchanged information, physical coupling, operating timescale, system scale, and embodiment. We separate results obtained with physical and simulated humanoids from those based on other robotic platforms or disembodied agents, because the latter do not fully capture humanoid-specific constraints such as balance, whole-body motion, contact, and physical interaction. We examine hardware, datasets, and benchmarks, and distinguish risks handled locally from those networking introduces, such as stale information or unsafe knowledge propagation. Finally, we identify open problems in control-aware networking, compact body-centric messages, physically grounded coordination, network- and embodiment-aware knowledge sharing, fleet-level social consistency, and accountable human oversight.

1 Introduction

Humanoids are moving toward connected fleets, but research has largely improved individual robots without explicit interfaces for shared information and decisions. This survey organizes networked humanoid intelligence around four functions and reviews them with humanoid-specific constraints in view.

1 Introduction

  • Connected humanoid fleets are emerging, while individual-robot research does not provide explicit interfaces for other robots to contribute information, coordinate decisions, or participate in joint control.Deployments may involve hundreds of robots at one factory site or larger populations across homes and workplaces.Source: §1 · p. 2

1.1 Why Humanoids Change the Networking Problem

  • Balance, whole-body control, physical contact, heterogeneous data, and embodiment differences make networking a control, interaction, and learning problem, not merely a communication layer.Delayed updates can disrupt synchronization and balance, while morphology and hardware differences constrain transfer of policies and experience.Source: §1.1 · pp. 2–3

1.2 Related Surveys and Positioning

  • Existing surveys separately emphasize individual humanoids, generic multi-robot systems, or network infrastructure, leaving inter-agent capabilities under humanoid-specific constraints unsystematized [16] [161] [36] [45] [48] [67] [133] [145].The survey positions itself as capability- and algorithm-oriented, organizing communication, coordination, collaboration, and social intelligence on humanoid platforms.Source: §1.2 · p. 3

1.3 Contributions and Organization

  • The survey organizes literature into four functions, compares humanoid and generic-robot evidence, assesses platforms and benchmarks, and distinguishes local from networking risks.It also derives open problems from the reviewed gaps.Source: §1.3 · p. 4

2 Foundations of Networked Humanoid Intelligence

The foundations frame networking around humanoids’ differing control timescales and define four functions by the questions they answer. A common factory scenario illustrates their distinct roles, while a taxonomy compares the literature across six dimensions and five embodiment levels.

2.1 The Humanoid as a Network Node

  • Balance requires fast local execution, while planning, learning, and model updates can tolerate edge-server communication; network dependence therefore varies by function.The survey contrasts a 10 Hz vision-language module and 120 Hz action module in GR00T N1 with Helix’s 200 Hz visuomotor policy and kilohertz joint control [12].Source: §2.1 · pp. 4–5

2.2 Four Functions, Four Questions

  • Communication asks what information reaches whom and when; coordination keeps decisions compatible, while cooperation specifies a shared goal and collaboration combines capabilities or experience.Social intelligence incorporates human intentions, expectations, responsibilities, and control; its demands also intersect with the other functions.Source: §2.2 · pp. 5–6

2.3 One Deployment, Four Functions: A Factory Scenario

  • The factory scenario separates status and event updates from control-timescale gait and force exchange, illustrating different communication demands within one fleet.Between shifts, larger skill updates tolerate looser delays.Source: §2.3 · p. 6
  • In the assembly scenario, fleet assignments respect capabilities and task constraints, while joint transport combines forces and shared experience requires adaptation across hand configurations.The worker-facing examples include intent inference, predictable motion, and traceable supervisor-approved plan deviations.Source: §2.3 · pp. 6–8

2.4 Taxonomy and Review Dimensions

  • The review taxonomy follows each function’s guiding question and analyzes interaction architecture, information, physical coupling, timescale, scale, and embodiment evidence.It distinguishes physical and simulated humanoids from other physical or simulated robots and abstract agents, whose results omit humanoid-specific constraints.Source: §2.4 · p. 8

3 Communication and Networking

Communication in humanoid fleets must match each function’s timing and information needs, while accounting for balance, heterogeneous traffic, and limited body-centric representations. The review covers connectivity, selective exchange, computation placement, and control-loop designs, and identifies gaps in physical-outcome evaluation.

3.1 Connectivity, Medium Access, and Middleware

  • Connectivity, medium access, and middleware offer infrastructure, scheduling, and discovery options, but direct humanoid-fleet evidence remains limited and mixed traffic is unevaluated.Only O-RACES among reviewed studies directly relates communication to physical outcomes such as humanoid falls.Source: §3.1 · p. 10

3.2 Selective and Learned Information Exchange

  • Selective exchange sends task-relevant scene content, event-triggered updates, or messages to selected senders or receivers instead of complete observations.Most approaches rely on abstract agents or idealized channels and have not been tested on humanoids or their contact-state and motion-intent messages.Source: §3.2 · p. 11

3.3 Communication-Computation Co-Design

  • 41% lower swarm energy consumption while meeting over 99% of task deadlines is reported for joint offloading and resource optimization.Under limited bandwidth, the method keeps more computation onboard; its workload-splitting assumption fits perception better than tightly coupled balance control.Source: §3.3 · p. 12

3.4 Communication Inside the Control Loop

  • Splitting whole-body control between edge computation and local feedback raises maximum tolerable constant delay from 40 to 90 ms for balancing and from 3 to 25 ms for walking.In simulations with a 37-DoF humanoid, the robot reuses the previous remote solution when communication is delayed.Source: §3.4 · p. 12
  • Up to 50% lower uplink latency and near-elimination of rough-terrain walking falls are reported for O-RACES proactive scheduling versus reactive scheduling.The evaluation used up to 16 simulated humanoids connected through a real Open RAN testbed.Source: §3.4 · p. 13

3.5 Summary and Lessons Learned

  • Communication requirements follow the consuming function: balance and physically coupled motion need fresh, bounded-delay information, while planning and model updates tolerate longer delays.For time-critical control, retransmitting an old packet may be worse than dropping it because it can misrepresent the current physical state.Source: §3.5 · p. 13
  • Humanoid fleets still lack compact body-centric messages for contact, gait phase, footsteps, interaction force, and predicted occupancy, including their validity duration.AoI and VoI capture why freshness and task relevance matter beyond message size.Source: §3.5 · p. 13
  • Communication, computation, and control placement should be co-designed, but existing joint-placement work mainly addresses delay-tolerant mapping and perception over stable links.Slow functions can be offloaded while time-critical feedback remains onboard to preserve control when remote results are delayed.Source: §3.5 · p. 13

4 Coordination

Coordination keeps humanoid decisions and actions mutually consistent through task allocation, shared plans, collision-free movement, and plan repair. Across these processes, humanoid coordination must account for changing physical constraints and information that can become outdated.

4 Coordination

  • Coordination links capability-based task assignment, shared roles and plans, collision-free movement, and plan repair when conflicts arise.The shared-workspace loop also accounts for body sway when reserving space.Source: §4 · pp. 13–14, Figure 4 · p. 14

4.1 Who Does What: Allocation and Scheduling

  • Capability-aware allocation ranges from scalar utilities and URDF-derived robot resumes to LLM task graphs and progress-based reassignment, yet often misses physical-state feasibility.SMART-LLM reaches 71% success versus 14% for rule-based allocation; DART-LLM raises construction success from 45% to 84%; COHERENT reaches 97.5% versus 75% without execution history.Source: §4.1 · pp. 14–15
  • In humanoid soccer, decentralized role reassignment takes 2.3 s versus 5.1 s with centralized coordination and maintains performance with more than 50% packet loss.The method considers distance to the ball, kicking ability, and battery level; another system assigns the attacker by predicted arrival time.Source: §4.1 · p. 15

4.2 Agreeing on Shared Decisions

  • Agreement mechanisms trade communication and computation cost against robustness and constraint handling; a hybrid planner achieves the highest success rate across four simulated tasks.Simultaneous wireless aggregation can agree on a formation reference using three orthogonal broadcast slots per update, regardless of team size.Source: §4.2 · pp. 15–16

4.3 Spatial Coordination

  • Spatial coordination combines shared collision-free regions, formation control, trajectory planning, and movement timing, but communication assumptions, deadlocks, and balance during stops constrain methods.Humanoid planning that accounts for center-of-mass sway improves narrow-gate success from 60% to 86.7% in simulation; another method adjusts footsteps in under 1 ms.Source: §4.3 · pp. 16–17

4.4 Runtime Conflict Resolution and Repair

  • Runtime repair ranges from pre-execution feasibility checks and failure diagnosis to leader waypoints for deadlocks and negotiated passing order at narrow passages.REFLECT’s recovery plans succeed in about 80% of simulated cases, but it focuses on single-robot tasks in static environments and is less effective at diagnosing low-level control failures.Source: §4.4 · pp. 17–18

4.5 Summary and Lessons Learned

  • Effective coordination pairs slow global proposals with fast local safety checks, but current checks rarely capture humanoid-specific constraints.Shared state should include a compact motion envelope for expected body occupancy, footsteps, allowable deviation, and information validity.Source: §4.5 · p. 18
  • Agreement cost should be weighed against decision validity: delay reduces tolerance, waiting may be infeasible, and outdated commitments call for targeted repair.Shared state should record each item’s source, scope, and validity period so only affected commitments need updating.Source: §4.5 · p. 18

5 Collaboration

Networked humanoid collaboration combines information, skills, experience, or physical capabilities; when robots share an object, their motions and balance become coupled. The section reviews physical joint execution, reasoning and capability composition, fleet learning, and remaining gaps in communication, body descriptions, and cost reporting.

5 Collaboration

  • Collaboration extends coordination: humanoids combine information, skills, experience, or physical capabilities, with shared-object contact coupling their balance.Studies describe information sharing [163], skill and experience sharing [86] [141], synchronized action, and leader-follower roles, compliant arms, and contact-state exchange during carrying [89].Source: §5 · pp. 18–19, Figure 5 · p. 19

5.1 Physically Coupled Joint Execution

  • Physical carrying uses centralized control or leader-follower feedback, but studies report stability constraints and leave network-condition effects unexplored.Two to four HRP-2s transported objects stably in simulation, while two Nao robots carried up to half a robot’s mass at about 2 cm/s; leader-follower approaches also showed oscillations or constrained motion.Source: §5.1 · pp. 19–20

5.2 Reasoning and Capability Composition

  • Reasoning and capability-composition studies combine agents’ knowledge, complementary roles, or predicted actions, but often use non-humanoid platforms or costly planning.CoELA reported 71% target-object delivery versus 43% for a single rule-based agent; COMBO’s repeated predictions make planning computationally expensive, and HetNet’s learned roles do not account fully for humanoid posture and balance.Source: §5.2 · pp. 20–21

5.3 Fleet Learning and Cross-Embodiment Transfer

  • Fleet learning shares experience, model updates, or cross-body policies, while communication costs, delays, and unsafe update propagation remain underexamined.Humanoid transfer studies include a shared locomotion policy adapted to an unseen body and a whole-body policy controlling seven physical humanoids without fine-tuning; deployed-fleet reliability and safety remain largely unexplored.Source: §5.3 · pp. 21–22

5.4 Summary and Lessons Learned

  • Physical cues and action inference can reduce messaging but cannot replace it; communication helps when a robot cannot observe or infer needed information [164].The section distinguishes control-loop information exchange during physical coupling from planning-timescale knowledge and skill combination, and experience sharing in fleet learning.Source: §5.4 · pp. 22–23
  • A usable description of contributing bodies remains lacking: simulator-derived partner state improved Harmanoid’s success, while embodiment descriptors sometimes reduced transfer [86] [89].The section proposes describing what a humanoid can safely contribute in its current state and how its skills map to other bodies.Source: §5.4 · p. 23
  • Studies report collaboration benefits but rarely measure its costs; only one physical-collaboration study describes its network, and Table 5 works do not measure communication cost [59] [141].Learning While Deploying sends a full VLA model to every robot after each update, but its communication overhead is unreported.Source: §5.4 · p. 23

6 Social Intelligence and Human Interaction

The section surveys how humanoids infer human intentions and expectations, coordinate responsibilities, and preserve meaningful human control. It finds that social state, fleet-level consistency, and timely oversight remain open challenges, with limited evidence from multi-humanoid settings.

6 Social Intelligence and Human Interaction

  • Humanoid interaction requires inferring human intentions, beliefs, and shared information, while accounting for trust, motion interpretation, proximity, and perceived safety.These concerns extend to fleets, where estimates of a person’s state must remain consistent across robots.Source: §6 · p. 23

6.1 Intentions: Social State Inference and Common Ground

  • Intent inference draws on probabilistic or neural models; common ground can be built through dialogue, nonverbal cues, shared mental models, gaze, gestures, and reused corrections.A robot can also adapt its plan using inferred human intentions, while another approach estimates what a person knows from objects they could observe.Source: §6.1 · pp. 23–24

6.2 Expectations: Legibility, Proximity, Norms, and Trust

  • Legible motion cut HERB goal-inference time by 33% without improving perceived safety; comfort evidence remains constrained by lab settings and motion-capture evaluation.Perceived safety also depends on robot appearance, movement, and proximity to people.Source: §6.2 · p. 24

6.3 Responsibilities and Control: Teaming, Teleoperation, and Oversight

  • Human-robot roles adapt to user preferences and workload, while control spans direct teleoperation through risk-prioritized fleet supervision.Fleet-DAgger prioritizes robots needing assistance but assumes immediate switching and does not consider network delays.Source: §6.3 · p. 25

6.4 Summary and Lessons Learned

  • Social-interaction evidence remains mostly single-robot; cross-robot state reconciliation and latency-aware, traceable human authority remain open fleet-level challenges.The reviewed work leaves unspecified how robots reconcile conflicting estimates of a person, while Fleet-DAgger assumes identical robots and does not account for network latency.Source: §6.4 · pp. 25–26

7 Evaluation, Safety, and Open Problems in Networked Humanoid Intelligence

The survey finds that available resources and evaluations cover pieces of networked humanoid intelligence, but not the integrated system. It separates local safety protections from network-added risks and outlines research directions spanning networking, physical coordination, knowledge sharing, and human oversight.

7.1 Platforms, Simulators, Datasets, and Benchmarks

  • No reviewed resource combines humanoid embodiment, multiple agents, network conditions, physical interaction, and human presence, leaving integrated evaluation unsupported.Table 6 assesses resources against these five properties; ✓ indicates provided, ◦ partial or optional, and ✗ absent.Source: §7.1 · p. 26, Table 6 · p. 27

7.2 What Current Evaluations Cannot Compare

  • Evaluation comparisons are constrained by incompatible communication-cost units, reliable-network assumptions, limited links to physical outcomes, and mostly one-to-one human studies.The reviewed work leaves fleet-level trust and supervision workload largely unmeasured.Source: §7.2 · p. 27
  • The survey calls for consistent reporting of communication volume and timing, varied network impairments, physical outcomes, fleet scaling, information usefulness, and human oversight measures.For physically coupled tasks, suggested outcomes include falls, balance-margin violations, and peak interaction forces.Source: §7.2 · p. 27

7.3 Safety: The Per-Robot Floor and the Risks the Network Adds

  • Network-dependent functions should fall back to local protection; above that per-robot safety floor, stale information and correlated communication can create fleet-level risks.Delay and packet loss can degrade control stability, and reactive medium access has caused simulated humanoid falls.Source: §7.3 · pp. 27–28
  • Fleet safety also requires containing propagation: erroneous or malicious updates and unsafe inter-robot messages can spread, while trust and approvals may extend across robots.The survey calls for freshness checks, local fallbacks, verification before fleet-wide deployment, and traceable origins for commitments, social estimates, and approvals.Source: §7.3 · p. 28

7.4 Open Problems and Research Agenda

  • Future networking should use gait- and contact-dependent deadlines, while compact body-centric messages could convey occupancy, footsteps, motion deviation, contact, and gait phase.A motion envelope could help robots assess whether a teammate’s plan remains physically compatible without exchanging complete trajectories.Source: §7.4 · p. 28
  • Coordination should validate assignments against current balance, posture, and load, while fleet learning should measure update costs, separate transferable components, and detect harmful updates.The survey proposes local whole-body-controller checks and gradual update deployment as starting points.Source: §7.4 · pp. 28–29
  • Shared social-state estimates should retain source, scope, uncertainty, and useful duration; oversight should allocate attention by learning needs and physical risk and record approval context.Across the four functions, the survey identifies freshness, embodiment, and human accountability as joint requirements.Source: §7.4 · p. 29

8 Conclusion

The conclusion reviews networked humanoid intelligence across communication, coordination, collaboration, and social interaction. It emphasizes humanoid-specific physical constraints and the need to jointly consider networking, planning, control, and human oversight.

8 Conclusion

  • The survey reviews four functions of networked humanoid intelligence: communication, coordination, collaboration, and social intelligence with human interaction.It examines how these functions support fleet operation and how humanoid physical characteristics shape their challenges.Source: §8 · p. 29
  • Balance, whole-body motion, physical contact, and hardware differences place additional demands on information exchange and joint decision-making.The conclusion notes that most studies treat the functions in isolation and give limited attention to network conditions, humanoid differences, or fleet safety.Source: §8 · p. 29
  • Future work should jointly consider communication, planning, physical control, and human oversight so fleets can work safely with people, adapt, and share knowledge.Source: §8 · p. 29

Related papers

Loading 2610.34740v1…