Pith. sign in

REVIEW 4 major objections 5 minor 55 references

SAGE: A Socially-Aware Generative Engine for Heterogeneous Multi-Agent Navigation

T0 review · 4 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read SAGE claims that a training-free safety-social energy term, injected into each step of a diffusion-based trajectory sampler, makes heterogeneous multi-robot teams measurably safer and more socially compliant while preserving task performanc

desk verdict SAGE is a well-engineered integration with a real circularity problem: its social-compliance metric is the same objective it optimizes, so the headline numbers are partly by construction. read the letter →

arxiv 2607.16619 v2 pith:GUCVB4PB submitted 2026-07-18 cs.RO cs.MA

Lan Hu , Minghui Liwang , Wenbo Zhu , Xinlei Yi , Yiguang Hong , Xianbin Wang , Zhenzhen Jiao , Seyyedali Hosseinalipour This is my paper
classification cs.ROcs.MA
keywords socially-awarenavigationmulti-robottrajectorypredictionandplanningheterogeneousmulti-agentsystemsdiffusiongenerativemodelsinference-timeenergyguidancepersonalspaceproxemicscollisionavoidancesafety-accuracytrade-offcontrol
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

SAGE asks whether a robot team navigating among pedestrians, cyclists, and vehicles can become safer and more socially compliant at inference time, without retraining its generative model. The paper's claim is yes: a differentiable safety-social energy that penalizes collisions, kinematic violations, missed waypoints, and intrusion into role-dependent personal-space ellipses is added to each denoising step of a conditional diffusion model, and its gradient corrects only the robot trajectories. Experiments on real pedestrian-trajectory scenes and synthetic multi-robot scenarios report consistent reductions in collision rate (about 20% on one benchmark), personal-space intrusion, and social-violation rate (16-27%), with only a few percent loss in trajectory accuracy and task metrics, and the mechanism scales to teams of 20 robots. If true, SAGE gives navigation a reusable dial for the safety-accuracy-task trade-off, unlike methods that fix such trade-offs during training and cannot be adjusted later.

What carries the argument

The load-bearing object is the safety-social energy function J of Eq. (28) together with the bounded robot-only gradient update of Eq. (33): at each reverse-diffusion step, the clean joint trajectory estimate is evaluated against J, its gradient is masked to robot components, clipped, scaled by a guidance strength λ, and applied as a bounded velocity correction before re-noising. The correction is deliberately masked so entity predictions stay untouched. A heterogeneous graph transformer—an attention encoder that learns separate projection matrices for each directed role pair such as robot-to-pedestrian versus pedestrian-to-robot—supplies the context that conditions the diffusion denoiser, c

What would settle it

An independent human-subject study in which participants rate the same guided trajectories for felt discomfort; if human ratings do not track SAGE's ellipse-derived SVR and PIR numbers, the social-compliance claim fails. A second check is a seeded reproduction with λ=0 versus λ>0 on identical scenes: if the reported collision-rate and social-violation-rate reductions cannot be reproduced, the guidance mechanism itself is not doing the work.

Watch

Extended reading notes

Core claim

On its own terms, SAGE's central discovery is that a learned conditional diffusion distribution over joint robot-entity trajectories can be steered at test time by a masked, clipped gradient of a composite energy J = w_col J_col + w_soc J_soc + w_kin J_kin + w_goal J_goal, applied only to the robot velocity components of the clean estimate before re-noising. The collision term repels robots from entities and from each other with a smooth potential; the social term uses anisotropic elliptical comfort zones whose longitudinal and lateral extents depend on the neighbor's role, with larger margins for pedestrians and smaller margins for vehicles, and only penalizes forward-half-plane intrusions;

Load-bearing premise

The claim stands on the assumption that social compliance is adequately captured by the fixed role-dependent elliptical comfort zones used both as the guidance objective and as the social-violation metric—and, on real pedestrian data, on the proxy protocol that treats a tracked pedestrian as a controllable robot whose waypoint is its own recorded future endpoint.

Editorial extensions

If this is right

  • Deploying the same trained model, a user can turn a single guidance-strength dial λ at inference time to trade safety and social compliance against trajectory accuracy, without retraining.
  • The mechanism scales to 20-robot teams: reported guided runtime stays below the 0.4 s control interval and Goal-FDE changes by less than 2% across team sizes, suggesting real-time receding-horizon operation is feasible.
  • Robot-to-robot collision guidance alone cuts inter-robot collision rate by roughly 10-11% and increases minimum inter-robot separation, while the graph's R2R edges complement guidance in preserving waypoint progress.
  • Jointly generating entity predictions and robot plans yields better goal-directed task performance than a two-stage pipeline that freezes entity predictions before planning, supporting the paper's coupling argument.
  • Additional weights such as the task-progress weight w_goal provide a second inference-time dial that trades goal-reaching against safety and social compliance in an interpretable, monotone way.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the ellipse-based social prior is portable, the same additive energy formulation could be repurposed for other soft constraints—legal clearance distances, noise limits, or culture-specific proxemics—by swapping σ_lon and σ_lat or adding new energy terms, all without retraining the diffusion model.
  • The ETH/UCY proxy protocol, which treats a tracked pedestrian as a controllable robot whose waypoint is its own recorded future endpoint, likely understates the difficulty of true goal-directed navigation; the reported real-world numbers should be re-verified in physical deployments where perception noise and reactive pedestrians change the safety-social balance.
  • Because the metric and the guidance objective use the same elliptical comfort regions, part of the measured SVR/PIR improvement may reflect the objective optimizing its own metric; independent human-judgment evaluation would separate genuine social acceptability from circularity.
  • The scaling result suggests a field-deployable architecture: at 20 robots the guidance is still within the control loop's time budget, so SAGE could be embedded in a receding-horizon planner for open environments with mixed traffic.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes SAGE, a framework for social multi-agent navigation among heterogeneous entities. It represents robots and surrounding agents with a directed heterogeneous graph processed by an HGT, uses a conditional diffusion model to jointly generate entity trajectory predictions and robot trajectory plans, and then applies a training-free energy guidance that iteratively corrects the robot-only component of the sampled velocities with respect to collision, kinematic, task-progress, and role-conditioned social-compliance terms. Experiments on ETH/UCY, SDD, and controlled simulations report reductions in collision rate, personal-space intrusion, social-violation rate, and guidance energy at modest ADE/FDE cost, together with scalability up to 20 robots and inference-time control of the safety-task trade-off.

Significance. If the empirical claims hold, SAGE provides a useful architectural contribution and a modular inference-time safety-social knob for heterogeneous multi-robot navigation. The HGT relation-specific attention and the joint prediction-planning diffusion formulation are reasonable, and the guidance mechanism is described in enough detail to be reproduced. The paper also contains a thorough ablation study (encoder choice, two-stage generation, guidance weights, pseudo-role construction, team size and density), which is a strength. However, the headline social-compliance result is not independent of the objective being minimized: SVR and PIR are computed from the same role-conditioned ellipses that the social energy term penalizes. The real-world 'navigation' evaluation on ETH/UCY is a proxy in which a tracked pedestrian is treated as a robot whose waypoint is the recorded future endpoint, which conflates prediction and planning. The significance is therefore conditional: the framework is promising, but the claimed validation of social acceptability, and to a lesser extent the quantitative safety claims, need re-framing or an external check.

major comments (4)
  1. [Sec. V-A, Eq. (11), Eq. (30), Tables I-II] The SVR metric is defined as the fraction of interactions inside the role-conditioned ellipse of Eq. (11), and J_soc in Eq. (30) is the same normalized-ellipse penetration with the same parameters. Reducing SVR by gradient descent on J_soc is therefore expected and does not validate the elliptical comfort model or establish social acceptability. The paper itself shows this in Table XI: Social Force achieves far lower SVR (0.0049 vs. 0.0185 controlled; 0.0043 vs. 0.0168 SocialGym-style) yet is not treated as more socially compliant. Please replace or supplement the social-compliance claims with an external measure (e.g., a human study, SocNavBench, or user study), or explicitly re-state the claims as 'reduces violations of the modeled social prior' rather than 'improves social compliance.'
  2. [Sec. V-A (ETH/UCY protocol)] On ETH/UCY, a tracked pedestrian is treated as a controllable robot and its recorded future endpoint is used as the waypoint. This turns the experiment into trajectory re-shaping: the 'robot' trajectories are human futures, and the reported CR/PIR/SVR improvements show how the guidance would alter a recorded human path, not how an actual robot planner performs in a shared environment. The same issue applies to the SDD single-robot proxy. The conclusions should be restricted accordingly, or an additional real-robot or realistic simulated-robot baseline should be added.
  3. [Tables I, II, V, VI, XII, XIII] The main tables report no error bars or statistical significance measures. Table II says values are averaged over three seeds, but variances are only shown in Fig. 2(b); Tables I and V report point estimates only. Many of the headline gains are small (e.g., SDD CR 0.1763 vs 0.1729, a 1.9% reduction; ETH/UCY CR 0.0441 vs 0.0353), so the claim that guidance 'consistently' improves safety and social compliance is not substantiated without variance and significance testing. Please report confidence intervals or per-seed/per-scene spreads and perform paired tests where appropriate.
  4. [Appendix A-A, Eq. (28), Eq. (33), Fig. 2(a)] The guidance mechanism has several free parameters: the scale lambda, the four weights (w_col, w_soc, w_kin, w_goal), sigma_col, w_R2R, and dataset-specific thresholds (d_safe, v_max, d_goal, etc.). Fig. 2(a) sweeps lambda, but the other weights are fixed without a sensitivity analysis. Since 'training-free controllability' is a central selling point, the paper should show that the main conclusions are stable over a reasonable range of hyperparameters or at least describe a principled procedure for choosing them. This would also address the concern that the results are partly driven by tuned constants.
minor comments (5)
  1. [Eqs. (7), (27), (33)] The notation S(out) is used both for the output set and for the diffusion state; Eq. (27) reuses S(out)_k' for a newly sampled noisy sequence. Please introduce distinct notation (e.g., tilde or caret) to avoid ambiguity.
  2. [Eq. (22)] The function rho_HGT is not defined precisely; please specify the order of residual connection, layer norm, and feed-forward update.
  3. [Sec. V-A (MD metric)] The definition of the minimum-separation (MD) metric is ambiguous: is it the minimum over time for each interaction, averaged over interactions? Please state the aggregation explicitly.
  4. [Table VI] The 'Count/Scene' column reports fractional values (e.g., 2.6130). This is understandable but should be defined in the caption as an average count per scene.
  5. [Sec. IV-C1, Eq. (29)] The collision potential phi_col is a sum of a Gaussian and a quadratic ReLU term. The Gaussian term is active at all distances and can dominate the gradient far from collision; consider clarifying the role of the Gaussian relative to the hard margin term.

Circularity Check

4 steps flagged · score 6.0 of 10

Guidance-related SVR/CR/Energy gains are mostly the optimizer descending on the same penalties later used as metrics; non-circular residue is the external accuracy trade-off.

  1. self definitional [Sec. IV-C1, Eq. (30) vs Sec. V-A SVR metric; ellipse values in Appendix A-A]
    "social-violation rate (SVR) metric measures the fraction within the role-conditioned ellipse in (11) ... J(soc)=Σ_t Σ_{r_j} Σ_{o_i} I[d(lon,t)≥0] exp(-||δ^t_{i,j}||_2^2) ... Following proxemics theory [17] and prior social-navigation literature [4], the role-dependent ... margins ... are set to (2.0,1.0) for PHAs, (2.6,1.2) for HNAAs, and (1.2,0.7) for SSAAs."

    Eq. (11) defines social compliance as staying outside a role-conditioned ellipse; Eq. (30) penalizes penetration of the same ellipse with the same σlon/σlat values, and Eq. (33) moves robot velocities along -∇J, which contains wsoc J(soc). Thus the reported SVR decrease is the optimizer lowering the exact quantity later counted as a violation. It does not validate the ellipse as a model of human comfort. The paper itself shows this tension: in Table XI Social Force has much lower SVR (0.0049 vs 0.0185) yet is not treated as more compliant, so SVR is not an independent social-validity measure.

  2. self definitional [Sec. IV-C1, Eq. (29) vs Sec. V-A CR metric]
    "collision rate (CR) metric measures the fraction of robot-entity and robot-robot interactions across the prediction horizon with separation below the collision margin ... J(col)=Σ_t[Σ_{r_j}(Σ_{o_i} φ_col(d^t_{j,i};m_{j,i}) + w_R2R Σ_{j<j'} φ_col(d^t_{j,j'};m_{j,j'}))]"

    J(col) in (29) uses the same safety margins m_{j,i}, m_{j,j'} as constraints (C1)-(C2), and CR counts pair separations below those same margins. Because Phase 3 descends ∇J including wcol J(col), the CR improvement is expected by construction. The correction is bounded and re-noised, so it is not fully trivial, but the CR column is not independent evidence that the robot is safer than what the objective already encodes.

2 more flagged steps
  1. self definitional [Sec. IV-C (Eqs. 28, 33) and Sec. V-A Energy metric]
    "Energy measures the total guidance objective J in (28), averaged over evaluation scenes and generated trajectory samples ... Compute the four energy terms following (28)-(32); Apply the bounded robot-only correction following (33)."

    The Energy metric is defined to be exactly J, the same objective whose negative gradient is used in the guided correction. Reporting lower Energy with guidance is therefore a direct check that the gradient step decreased its own loss, not an external validation. The meaningful external outcomes are the R-ADE/Goal-FDE trade-offs and the parts of safety not perfectly encoded in J.

  2. self definitional [Sec. IV-C1, Eq. (32) and Table IX]
    "the task-progress energy is defined as follows: J(goal)=Σ_{r_j} ||p(r,T(prd))_j − g_j||_2^2 ... As wgoal increases from 0 to 2.0, R-ADE decreases from 1.4740 to 0.8267, Goal-FDE decreases from 2.4349 to 1.1677, and Goal-SR increases from 0.2120 to 0.3169"

    Goal-FDE is defined as the terminal distance to the assigned waypoint, which is exactly J(goal) in Eq. (32). Sweeping wgoal upward strengthens the gradient on this same term, so the Goal-FDE and Goal-SR improvements in Table IX are the optimizer succeeding on its own objective. The non-circular information in that sweep is the accompanying increase in R-ADE, CR, PIR, SVR, and Energy.

full rationale

The diffusion model and scene encoder are externally checkable: R-ADE, Goal-FDE, and E-FDE are compared against recorded ground-truth trajectories and against independent baselines (Trajectron++, MID, Constant Velocity), and the unguided-vs-guided comparison holds the trained model fixed. Those parts are not circular. However, the paper's central safety/social-compliance claims rest mainly on SVR, CR, and Energy columns, and each of those is the same quantity the guidance is designed to minimize: SVR counts points inside the ellipse used in J_soc, CR counts violations of the same margins used in J_col, and Energy is J itself. Goal-FDE is likewise the exact J(goal) term in the w_goal sweep. Therefore the statement that guidance consistently improves safety and social compliance is only partially supported; the social-compliance part especially lacks an external validity check, a point reinforced by Table XI where Social Force achieves lower SVR but is not judged more compliant. No load-bearing self-citation chain or imported uniqueness theorem was found. Score 6 reflects partial, not complete, circularity: the core safety-social claims are partly by construction, while the accuracy/task trade-off analysis retains independent content.

Assumptions & free parameters 5 free parameters · 4 assumptions · 1 invented entities

The paper introduces no new physical entity, but its claimed social-compliance gains rest on hand-set social and kinematic thresholds and on metrics aligned with the energy being minimized. The list above captures the main hand-chosen quantities and domain assumptions that the central claims require.

free parameters (5)
  • guidance scale λ = 0.05 (ETH/UCY); 0.03 in pseudo-role ablations
    Controls the strength of the energy-gradient correction; swept on the same evaluation scenes in Fig. 2(a), with no held-out tuning split reported.
  • guidance weights (wcol, wsoc, wkin, wgoal) = (2.0, 1.0, 0.1, 0.3)
    Hand-set weights balancing collision, social, kinematic, and goal energies; central to the behavior of Phase 3.
  • role-dependent social margins (σlon, σlat) = PHA (2.0,1.0), HNAA (2.6,1.2), SSAA (1.2,0.7)
    Hand-set from proxemics; they define both the Eq. (11) constraint and the SVR metric, causing metric-objective circularity.
  • collision potential scale σcol and robot-robot weight wR2R = σcol=0.8; wR2R=1.5 controlled, 1.0 SDD
    Shape the anti-collision potential in Eq. (29); chosen by hand.
  • safety and kinematic thresholds (d_safe, vmax, vlat_max, ωmax, d_goal) = ETH/UCY (0.45,2.5,0.15,1.2,0.5); controlled (0.30,3.0,0.15,1.2,0.5); SDD (0.35,3.0,0.15,1.2,0.6)
    Define CR, Sp-VR, and goal-success metrics; set once per dataset, not fitted, but load-bearing for the reported violation rates.
assumptions (4)
  • domain assumption Hall's proxemics and fixed role-dependent elliptical comfort regions capture socially compliant navigation.
    Introduced in Sec. III-C1/C4 and App. A-A; the central SVR/PIR metrics and J_soc energy are built from these hand-set margins.
  • domain assumption A target pedestrian in ETH/UCY, with its own recorded future endpoint as waypoint, is a valid proxy for a controllable robot.
    Sec. V-A defines the proxy-navigation protocol; all ETH/UCY safety and social claims rely on it.
  • domain assumption Gradient-based energy correction during reverse diffusion keeps samples close enough to the learned joint distribution pθ that accuracy metrics remain meaningful.
    Eq. (33) modifies clean estimates with a clipped gradient; no formal or systematic distribution-shift analysis is provided.
  • domain assumption The four role categories PHA/HNAA/SSAA/EAA and ordered relation IDs suffice to represent asymmetric multi-agent interactions.
    Used throughout Phase 1; cannot be validated on ETH/UCY, which lacks semantic labels and requires pseudo-role construction (Table X).
invented entities (1)
  • Role-conditioned heterogeneous anisotropic social potential fields (elliptical comfort regions)
    purpose: Represent soft social norms and guide robot trajectories at inference
    A modeling construct defined by hand-set σlon/σlat; no independent human-comfort validation exists, and SVR is measured from the same field, creating circularity.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SAGE: A Socially-Aware Generative Engine for Heterogeneous Multi-Agent Navigation." pith.science (2026). https://pith.science/paper/GUCVB4PB

@misc{pith2026260716619,
  author       = {Pith},
  title        = {Pith review of: SAGE: A Socially-Aware Generative Engine for Heterogeneous Multi-Agent Navigation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GUCVB4PB}},
  note         = {Machine review of arXiv:2607.16619}
}
read the original abstract

Safe and socially compliant navigation in open human-robot environments requires robots to reason about heterogeneous participants with different dynamics, autonomy levels, and social roles. Existing trajectory prediction and planning methods often rely on homogeneous interaction assumptions or enforce only geometric collision constraints, making it difficult to jointly model asymmetric interactions, coupled prediction-planning, and soft social norms. This paper proposes SAGE, a socially-aware generative engine for heterogeneous multi-agent navigation. SAGE represents robots and surrounding entities as a directed heterogeneous graph and employs a Heterogeneous Graph Transformer (HGT) to encode type-specific asymmetric interactions. Conditioned on the resulting context, a diffusion-based generative module jointly models future entity trajectories and robot trajectory plans. During inference, a training-free safety-social energy guidance mechanism refines sampled robot trajectories using differentiable collision, kinematic, task-progress, and role-conditioned social-compliance terms. Extensive experiments on real-world (ETH/UCY and SDD) and synthetic datasets verify the effectiveness of SAGE in improving safety and social compliance while maintaining task performance. The proposed guidance mechanism consistently reduces collision and social-violation rates, scales to teams of up to 20 robots, and enables explicit control of the safety-accuracy-task trade-off without retraining. These findings demonstrate the potential of SAGE as a scalable framework for socially-aware multi-agent navigation in complex environments.

Figures

Figures reproduced from arXiv: 2607.16619 by the authors.

Figure 1
Figure 1. Overview of the three phases of SAGE. Phase 1 encodes scene [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Guidance effects across evaluation regimes. (a) Safety, social [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 4
Figure 4. Qualitative comparisons in SDD scenes, where [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗
Figures from the paper (2 more)
Figure 3
Figure 3. Figure 3: Controlled cases: bottleneck yielding, multi-robot crossing, vehicle [PITH_FULL_IMAGE:figures/full_fig_p011_3.png]
Figure 5
Figure 5. Figure 5: Detailed illustration of one safety-social guidance step during reverse [PITH_FULL_IMAGE:figures/full_fig_p014_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

55 extracted references · 2 linked inside Pith

  1. [1]

    Deep learning in robotics: Survey on model structures and training strategies,

    A. I. Karoly, P. Galambos, J. Kuti, and I. J. Rudas, “Deep learning in robotics: Survey on model structures and training strategies,”IEEE Trans. Syst., Man, Cybern., Syst., vol. 51, no. 1, pp. 266–279, Jan. 2021

    2021

  2. [2]

    Socially compliant mobile robot navigation via inverse reinforcement learning,

    H. Kretzschmar, M. Spies, C. Sprunk, and W. Burgard, “Socially compliant mobile robot navigation via inverse reinforcement learning,” Int. J. Robot. Res., vol. 35, no. 11, pp. 1289–1307, Sep. 2016

    2016

  3. [3]

    Interactive model predictive control for robot navigation in dense crowds,

    Y . Chen, F. Zhao, and Y . Lou, “Interactive model predictive control for robot navigation in dense crowds,”IEEE Trans. Syst., Man, Cybern., Syst., vol. 52, no. 4, pp. 2289–2301, Apr. 2022

    2022

  4. [4]

    Social force model for pedestrian dynamics,

    D. Helbing and P. Molnar, “Social force model for pedestrian dynamics,” Phys. Rev. E, vol. 51, no. 5, pp. 4282–4286, May 1995

    1995

  5. [5]

    Social behavior for autonomous vehicles,

    W. Schwarting, A. Pierson, J. Alonso-Mora, S. Karaman, and D. Rus, “Social behavior for autonomous vehicles,”Proc. Natl. Acad. Sci. USA, vol. 116, no. 50, pp. 24 972–24 978, Dec. 2019

    2019

  6. [6]

    Core challenges of social robot navigation: A survey,

    C. Mavrogianniset al., “Core challenges of social robot navigation: A survey,”ACM Trans. Hum.-Robot Interact., vol. 12, no. 3, pp. 1–39, Sep. 2023

    2023

  7. [7]

    Social GAN: Socially acceptable trajectories with generative adversarial networks,

    A. Gupta, J. Johnson, L. Fei-Fei, S. Savarese, and A. Alahi, “Social GAN: Socially acceptable trajectories with generative adversarial networks,” inProc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), 2018, pp. 2255–2264

    2018

  8. [8]

    Social LSTM: Human trajectory prediction in crowded spaces,

    A. Alahi, K. Goel, V . Ramanathan, A. Robicquet, L. Fei-Fei, and S. Savarese, “Social LSTM: Human trajectory prediction in crowded spaces,” inProc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), 2016, pp. 961–971

    2016

Show all 55 references
  1. [9]

    Social-STGCNN: A social spatio-temporal graph convolutional neural network for human trajectory prediction,

    A. Mohamed, K. Qian, M. Elhoseiny, and C. Claudel, “Social-STGCNN: A social spatio-temporal graph convolutional neural network for human trajectory prediction,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2020, pp. 14 424–14 432

    2020

  2. [10]

    Reciprocal n-body collision avoidance,

    J. Van Den Berg, S. J. Guy, M. Lin, and D. Manocha, “Reciprocal n-body collision avoidance,” inProc. 14th Int. Symp. Robot. Res. (ISRR), 2011, pp. 3–19

    2011

  3. [11]

    Collision and deadlock avoidance in multirobot systems: A distributed approach,

    Y . Zhou, H. Hu, Y . Liu, and Z. Ding, “Collision and deadlock avoidance in multirobot systems: A distributed approach,”IEEE Trans. Syst., Man, Cybern., Syst., vol. 47, no. 7, pp. 1712–1726, Jul. 2017

    2017

  4. [12]

    Human motion trajectory prediction: A survey,

    A. Rudenko, L. Palmieri, M. Herman, K. M. Kitani, D. M. Gavrila, and K. O. Arras, “Human motion trajectory prediction: A survey,”Int. J. Robot. Res., vol. 39, no. 8, pp. 895–935, Jul. 2020

    2020

  5. [13]

    Game-based approximate optimal motion planning for safe human-swarm interaction,

    M. Li, J. Qin, J. Li, Q. Liu, Y . Shi, and Y . Kang, “Game-based approximate optimal motion planning for safe human-swarm interaction,” IEEE Trans. Cybern., vol. 54, no. 10, pp. 5649–5660, Oct. 2024

    2024

  6. [14]

    AgentFormer: Agent- aware transformers for socio-temporal multi-agent forecasting,

    Y . Yuan, X. Weng, Y . Ou, and K. M. Kitani, “AgentFormer: Agent- aware transformers for socio-temporal multi-agent forecasting,” inProc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), 2021, pp. 9813–9823

    2021

  7. [15]

    Social-BiGAT: Multimodal trajectory forecasting using Bicycle-GAN and graph attention networks,

    V . Kosaraju, A. Sadeghian, R. Martín-Martín, I. Reid, H. Rezatofighi, and S. Savarese, “Social-BiGAT: Multimodal trajectory forecasting using Bicycle-GAN and graph attention networks,” inAdv. Neural Inf. Process. Syst., vol. 32, 2019, pp. 137–146

    2019

  8. [16]

    Heterogeneous graph transformer,

    Z. Hu, Y . Dong, K. Wang, and Y . Sun, “Heterogeneous graph transformer,” inProc. Web Conf. (WWW), 2020, pp. 2704–2710

    2020

  9. [17]

    E. T. Hall,The Hidden Dimension. Garden City, NY , USA: Doubleday, 1966

    1966

  10. [18]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” inAdv. Neural Inf. Process. Syst., vol. 33, 2020, pp. 6840–6851

    2020

  11. [19]

    Stochastic trajectory prediction via motion indeterminacy diffusion,

    T. Guet al., “Stochastic trajectory prediction via motion indeterminacy diffusion,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2022, pp. 17 113–17 122

    2022

  12. [20]

    Score-based generative modeling through stochastic differential equations,

    Y . Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole, “Score-based generative modeling through stochastic differential equations,” inProc. Int. Conf. Learn. Represent. (ICLR), 2021

    2021

  13. [21]

    Trajectron++: Dynamically-feasible trajectory forecasting with heterogeneous data,

    T. Salzmann, B. Ivanovic, P. Chakravarty, and M. Pavone, “Trajectron++: Dynamically-feasible trajectory forecasting with heterogeneous data,” in Proc. Eur. Conf. Comput. Vis. (ECCV), 2020, pp. 683–700

    2020

  14. [22]

    Structural properties and classification of kinematic and dynamic models of wheeled mobile robots,

    G. Campion, G. Bastin, and B. Dandrea-Novel, “Structural properties and classification of kinematic and dynamic models of wheeled mobile robots,”IEEE Trans. Robot. Autom., vol. 12, no. 1, pp. 47–62, Feb. 1996

    1996

  15. [23]

    SoPhie: An attentive GAN for predicting paths compliant to social and physical constraints,

    A. Sadeghian, V . Kosaraju, A. Sadeghian, N. Hirose, H. Rezatofighi, and S. Savarese, “SoPhie: An attentive GAN for predicting paths compliant to social and physical constraints,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2019, pp. 1349–1358

    2019

  16. [24]

    Hetero- geneous graph neural network,

    C. Zhang, D. Song, C. Huang, A. Swami, and N. V . Chawla, “Hetero- geneous graph neural network,” inProc. 25th ACM SIGKDD Int. Conf. Knowl. Discov. Data Min., 2019, pp. 793–803

    2019

  17. [25]

    GMP: A genetic mission planner for heterogeneous multirobot system applications,

    B. Miloradovic, B. Curuklu, M. Ekstrom, and A. V . Papadopoulos, “GMP: A genetic mission planner for heterogeneous multirobot system applications,”IEEE Trans. Cybern., vol. 52, no. 10, pp. 10 627–10 638, Oct. 2022

    2022

  18. [26]

    Optimal task and motion planning and execution for multiagent systems in dynamic environments,

    M. Faroni, A. Umbrico, M. Beschi, A. Orlandini, A. Cesta, and N. Pedrocchi, “Optimal task and motion planning and execution for multiagent systems in dynamic environments,”IEEE Trans. Cybern., vol. 54, no. 6, pp. 3366–3377, Jun. 2024

    2024

  19. [27]

    Deep unsupervised learning using nonequilibrium thermodynamics,

    J. Sohl-Dickstein, E. Weiss, N. Maheswaranathan, and S. Ganguli, “Deep unsupervised learning using nonequilibrium thermodynamics,” inProc. Int. Conf. Mach. Learn. (ICML), 2015, pp. 2256–2265

    2015

  20. [28]

    Planning with diffusion for flexible behavior synthesis,

    M. Janner, Y . Du, J. B. Tenenbaum, and S. Levine, “Planning with diffusion for flexible behavior synthesis,” inProc. 39th Int. Conf. Mach. Learn. (ICML), 2022, pp. 9902–9915

    2022

  21. [29]

    Is conditional generative modeling all you need for decision-making?

    A. Ajay, Y . Du, A. Gupta, J. Tenenbaum, T. Jaakkola, and P. Agrawal, “Is conditional generative modeling all you need for decision-making?” inProc. Int. Conf. Learn. Represent. (ICLR), 2023

    2023

  22. [30]

    MotionDiffuser: Controllable multi-agent motion prediction using diffu- sion,

    C. M. Jiang, A. Cornman, C. Park, B. Sapp, Y . Zhou, and D. Anguelov, “MotionDiffuser: Controllable multi-agent motion prediction using diffu- sion,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2023, pp. 9644–9653

    2023

  23. [31]

    SafeDiffuser: Safe planning with diffusion probabilistic models,

    W. Xiao, T.-H. Wang, C. Gan, R. Hasani, M. Lechner, and D. Rus, “SafeDiffuser: Safe planning with diffusion probabilistic models,” inProc. Int. Conf. Learn. Represent. (ICLR), 2025

    2025

  24. [32]

    Guided conditional diffusion for controllable traffic simulation,

    Z. Zhonget al., “Guided conditional diffusion for controllable traffic simulation,” inProc. IEEE Int. Conf. Robot. Autom. (ICRA), 2023, pp. 3560–3566

    2023

  25. [33]

    A diffusion-model of joint interactive navigation,

    M. Niedobaet al., “A diffusion-model of joint interactive navigation,” inAdv. Neural Inf. Process. Syst., vol. 36, 2023, pp. 27 440–27 459

    2023

  26. [34]

    JPPD: Joint prediction–planning diffusion with differ- entiable safety guidance for dynamic obstacle avoidance in intelligent transportation systems,

    J. Wu and S. Yu, “JPPD: Joint prediction–planning diffusion with differ- entiable safety guidance for dynamic obstacle avoidance in intelligent transportation systems,”arXiv preprint arXiv:2606.20686, Jun. 2026

    2026 arXiv

  27. [35]

    GALC: Guided amplified learning with Lipschitz constraint for robust trajectory gen- eration,

    Z. Lin, Z. C. Chen, G. Zhu, J. Chen, and J. Li, “GALC: Guided amplified learning with Lipschitz constraint for robust trajectory gen- eration,”IEEE Trans. Cybern., pp. 1–11, Mar. 2026, early access, doi: 10.1109/TCYB.2026.3668987

    2026

  28. [36]

    Human-aware robot navigation: A survey,

    T. Kruse, A. K. Pandey, R. Alami, and A. Kirsch, “Human-aware robot navigation: A survey,”Robot. Auton. Syst., vol. 61, no. 12, pp. 1726–1743, Dec. 2013

    2013

  29. [37]

    Crowd-robot interaction: Crowd-aware robot navigation with attention-based deep reinforcement learning,

    C. Chen, Y . Liu, S. Kreiss, and A. Alahi, “Crowd-robot interaction: Crowd-aware robot navigation with attention-based deep reinforcement learning,” inProc. IEEE Int. Conf. Robot. Autom. (ICRA), 2019, pp. 6015–6022

    2019

  30. [38]

    Social-Transmotion: Promptable human trajectory prediction,

    S. Saadatnejad, Y . Gao, K. Messaoud, and A. Alahi, “Social-Transmotion: Promptable human trajectory prediction,” arXiv:2312.16168, 2023. [Online]. Available: https://arxiv.org/abs/2312.16168

    2023 arXiv

  31. [39]

    Evaluation of socially-aware robot navigation,

    Y . Gao and C.-M. Huang, “Evaluation of socially-aware robot navigation,” Front. Robot. AI, vol. 8, p. 721317, Jan. 2022

    2022

  32. [40]

    Soc- NavBench: A grounded simulation testing framework for evaluating social navigation,

    A. Biswas, A. Wang, G. Silvera, A. Steinfeld, and H. Admoni, “Soc- NavBench: A grounded simulation testing framework for evaluating social navigation,”ACM Trans. Hum.-Robot Interact., vol. 11, no. 3, pp. 1–24, Sep. 2022

    2022

  33. [41]

    Are socially-aware trajectory prediction models really socially-aware?

    S. Saadatnejad, M. Bahari, P. Khorsandi, M. Saneian, S.-M. Moosavi- Dezfooli, and A. Alahi, “Are socially-aware trajectory prediction models really socially-aware?”Transp. Res. C, Emerg. Technol., vol. 141, p. 103705, Aug. 2022

    2022

  34. [42]

    Coordination and control of complex network systems with switching topologies: A survey,

    G. Wen, X. Yu, W. Yu, and J. Lu, “Coordination and control of complex network systems with switching topologies: A survey,”IEEE Trans. Syst., Man, Cybern., Syst., vol. 51, no. 10, pp. 6342–6357, Oct. 2021

    2021

  35. [43]

    Distributed UA V swarm formation and collision avoidance strategies over fixed and switching topologies,

    J. Wu, C. Luo, Y . Luo, and K. Li, “Distributed UA V swarm formation and collision avoidance strategies over fixed and switching topologies,” IEEE Trans. Cybern., vol. 52, no. 10, pp. 10 969–10 979, Oct. 2022

    2022

  36. [44]

    Distributed task rescheduling with time constraints for the optimization of total task allocations in a multirobot system,

    J. Turner, Q. Meng, G. Schaefer, A. Whitbrook, and A. Soltoggio, “Distributed task rescheduling with time constraints for the optimization of total task allocations in a multirobot system,”IEEE Trans. Cybern., vol. 48, no. 9, pp. 2583–2597, Sep. 2018

    2018

  37. [45]

    Learning-based multi-UA V flocking control with limited visual field and instinctive repulsion,

    C. Bai, Y . Peng, H. Piao, W. Pan, and J. Guo, “Learning-based multi-UA V flocking control with limited visual field and instinctive repulsion,”IEEE Trans. Cybern., vol. 54, no. 1, pp. 462–475, Jan. 2024

    2024

  38. [46]

    Graph attention networks,

    P. Veli ˇckovi´c, G. Cucurull, A. Casanova, A. Romero, P. Liò, and Y . Bengio, “Graph attention networks,” inProc. Int. Conf. Learn. Represent. (ICLR), 2018

    2018

  39. [47]

    Diffusion models beat GANs on image synthesis,

    P. Dhariwal and A. Nichol, “Diffusion models beat GANs on image synthesis,” inAdv. Neural Inf. Process. Syst., vol. 34, 2021, pp. 8780– 8794

    2021

  40. [48]

    Learning social etiquette: Human trajectory understanding in crowded scenes,

    A. Robicquet, A. Sadeghian, A. Alahi, and S. Savarese, “Learning social etiquette: Human trajectory understanding in crowded scenes,” inProc. Eur. Conf. Comput. Vis. (ECCV), 2016, pp. 549–565. 13 APPENDIXA ADDITIONALEXPERIMENTALDETAILS ANDRESULTS This appendix collects impleme...

    2016

  41. [49]

    Heterogeneous encoder:To assess the contribution of heterogeneous interaction modeling, we replace the HGT with a homogeneous graph-attention encoder that uses a single shared relation type and omits role embeddings, thereby removing relation-specific attention projections. As...

    1915

  42. [50]

    Joint vs. two-stage generation:To evaluate the benefit of jointly generating robot and surrounding-entity trajectories, Table VIII compares SAGE with a two-stage variant in which entity trajectories are first generated and then fixed while robot trajectories are subsequently g...

  43. [51]

    As wgoal increases from 0 to 2.0, R-ADE decreases from 1.4740 to 0.8267, TABLE VII HGTVERSUS HOMOGENEOUS GRAPH-ATTENTION ENCODING ONSDD UNDER GUIDED SAMPLING

    Task-progress recovery via wgoal:To characterize the trade-off between goal-directed task performance and safety/social compliance, Table IX sweeps the task-progress weight wgoal on SDD under guided sampling. As wgoal increases from 0 to 2.0, R-ADE decreases from 1.4740 to 0.8...

    1951

  44. [52]

    Pseudo-role construction robustness:Since ETH/UCY does not provide semantic agent labels, Table X evaluates whether the effectiveness of guidance depends on the pseudo- role construction strategy. We consider three alternatives: speed_threshold, which assigns roles using fixed...

  45. [53]

    In the controlled simulation, Social Force achieves lower R2E-CR (0.0136 vs

    Social Force comparison:Table XI compares SAGE with the deterministic Social Force planner [4], which uses hand- crafted attractive and repulsive potentials to regulate robot motion. In the controlled simulation, Social Force achieves lower R2E-CR (0.0136 vs. 0.0371), R2R-CR (...

  46. [54]

    SDD multi-robot setting:Table XII evaluates the SAGE pipeline in multi-robot navigation settings constructed from SDD scenes using the official semantic annotations. As the number of robots increases from Nr = 3 to 12, Goal-FDE increases from 0.1260 to 0.3254 and Goal-SR decre...

  47. [55]

    SocialGym-style scale and density:To further stress-test SAGE beyond the primary controlled-simulation setting, we evaluate whether the benefits of guidance persist across a broader range of structured and crowded navigation scenarios and under varying robot-team sizes and ent...

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.