SoGuDiff: Socially Guided Diffusion for Steerable, Norm-Grounded Robot Navigation

Christian Schaible; Haoran Ji; Yash Vardhan Pant; Stephen L. Smith

cs.ROSep 24, 2026

SoGuDiff: Socially Guided Diffusion for Steerable, Norm-Grounded Robot Navigation

Authors: Christian Schaible, Haoran Ji, Yash Vardhan Pant, Stephen L. Smith

Organizations: Department of Electrical and Computer Engineering, University of Waterloo, Waterloo, Canada · Department of Computing and Software, McMaster University, Hamilton, Canada

Abstract

Beyond collision avoidance, socially competent robot navigation requires adherence to implicit social conventions that vary across contexts, cultures, and deployment requirements. Many conventional navigation policies learn a single normative behavior, either through reinforcement learning against a fixed reward function or imitation of human demonstrations, exposing no interface for adjusting that conduct at runtime. We present a diffusion-based navigation framework whose social behavior can be tuned at deployment: a desired style is specified, such as how closely the robot passes, which side it yields to, or how much it defers to groups, and the planner adapts accordingly. Continuous style axes can be followed independently or composed, spanning a behavioral space rather than discrete, primitive-based specifications. A feasibility projection layer separates learned social behavior from kinematic feasibility and collision avoidance. A single-axis sweep illustrates a tradeoff curve that strictly dominates the evaluated fixed-behavior baseline configurations, and stylistic differences are replicated in real-world demonstrations.

Figures & tables

Fig. 1 : SoGuDiff steers a single diffusion planner along interpretable, composable social style axes at deployment. For each style, candidate trajectories are sampled, one is selected and feasibility projection ensues. Varying the style vector, without retraining, yields controllable, compositional behaviors. Human meshes are from HuNavSim [ 11 ] .

Fig. 2 : Overview of SoGuDiff. (a) Training: scene inputs (initial state, goal, pedestrian states, and occupancy map) and the social style vector are embedded as tokens, with structured conditioning dropout over the scene and style tokens. A one-dimensional conditional U-Net is trained to denoise trajectories. (b) Inference: the network is queried under decomposed conditions (unconditional, scene, and scene-plus-each-style-axis) and combined by per-axis classifier-free guidance to steer and compose style. Candidate trajectories are denoised in parallel, the best is selected, and feasibility projection either produces feasible actions or triggers a controlled stop if acceptance criteria are not met.

Term	Convention captured	Axis
$g_{\mathrm{prox}}$	keep clear of pedestrian space	$s_{\mathrm{prox}}$
$g_{\mathrm{rear}}$	avoid approaching from directly behind	$s_{\mathrm{prox}}$
$g_{\mathrm{side}}$	pass on the convention side	$s_{\mathrm{pass}}$
$g_{\mathrm{ttc}}$	anticipate and defer to oncomers	$s_{\mathrm{yield}}$
$g_{\mathrm{cut}}$	do not cut in front of a pedestrian	$s_{\mathrm{yield}}$
$g_{\mathrm{group}}$	do not split social groups	$s_{\mathrm{group}}$

TABLE I : Social cost terms and the style axis governing each.

Method	SR/TR/CR	TTG	PL	MeanD	PSI
Method	$\uparrow$ (%)	$\downarrow$ (s)	$\downarrow$ (m)	$\uparrow$ (m)	$\downarrow$ (%)
SFM [ 3 ]	84/ 0 1/15	11.12	0 9.27	3.49	11.25
ORCA [ 4 ]	98/ 0 0/ 0 2	10.75	0 9.30	3.50	11.28
CADRL [ 5 ]	97/ 0 2/ 0 1	11.17	10.12	3.66	0 4.78
LSTM-RL [ 43 ]	94/ 0 3/ 0 3	11.56	10.81	3.67	0 5.20
SARL [ 6 ]	96/ 0 2/ 0 2	11.40	10.32	3.71	0 5.95

TABLE II : Success, efficiency and proximity metrics for SoGuDiff and baselines on a random set of 500 crowd scenes. For each metric, the best value is bolded and the second-best is underlined. Holonomically trained baselines are denoted by † .

Fig. 3 : The TTG vs PSI tradeoff for fixed-behavior baselines and SoGuDiff, traced by varying proximity style, over 500 random crowd scenes. Holonomically trained baselines are denoted by † .

Panel of Figure 6 — Fig. 4 : Closed-loop trajectory distributions for the nondeterministic SoGuDiff planner using swept single-axis styles across 10 randomly seeded runs on representative scenes. Darker trajectories correspond to $-1$ styles, with lighter corresponding to $+1$ for the (a) proximity, (b) passing-side, (c) yielding and (d) group deference style axis sweeps.

Social Style	SR/TR/CR	MeanD	PSB	TCR	PCR	GSR
Social Style	$\uparrow$ (%)	$\uparrow$ (m)	$\updownarrow$ (-)	$\downarrow$ (%)	$\downarrow$ (%)	$\downarrow$ (%)
Neutral	0 96/ 0 4/ 0 0	2.38	+ 0.071	0.38	2.31	17.97
prox + 1	0 95/ 0 5/ 0 0	3.28	+ 0.232	0.02	0.00	0 0.00
prox + 0.5	0 91/ 0 9/ 0 0	2.75	+ 0.181	0.03	0.01	0 0.55
prox - 0.5	100/ 0 0/ 0 0	2.29	- 0.111	1.94	2.93	21.12
prox - 1	100/ 0 0/ 0 0	2.27	- 0.148	2.37	3.40	20.32

TABLE III : SoGuDiff behavior on a set of 500 geometric scenarios, using swept single-axis styles. For each swept axis, values pertaining to the style-specific metrics are shaded.

Fig. 5 : Closed-loop trajectory distributions for the SoGuDiff planner using composed axis styles across 10 randomly seeded runs each.

Social Style	SR/TR/CR	MeanD	PSB	TCR	PCR	GSR
Social Style	$\uparrow$ (%)	$\uparrow$ (m)	$\updownarrow$ (-)	$\downarrow$ (%)	$\downarrow$ (%)	$\downarrow$ (%)
$[$ -1 , 0 , 0 , +1 $]$	100/ 0 0/ 0 0	2.29	- 0.081	2.00	2.79	17.23
$[$ 0 , +1 , +1 , 0 $]$	100/ 0 0/ 0 0	2.48	+ 0.647	0.27	0.61	0 8.94
$[$ +1 , -1 , -1 , 0 $]$	100/ 0 0/ 0 0	2.49	- 0.249	0.57	1.46	0 7.49
$[$ 0 , -1 , +1 , -1 $]$	0 98/ 0 2/ 0 0	2.45	- 0.412	0.40	1.01	11.37
$[$ +1 , +1 , +1 , +1 $]$	100/ 0 0/ 0 0	2.87	+ 0.405	0.06	0.18	0 0.97

TABLE IV : SoGuDiff behavior on a set of 500 geometric scenarios, using composed axis styles. Social styles follow the

[s_{\mathrm{prox}},s_{\mathrm{pass}},s_{\mathrm{yield}},s_{\mathrm{group}}]

notation.

Style, Ablation	SR/TR/CR	MeanD	PSB	TCR	PCR	GSR
Style, Ablation	$\uparrow$ (%)	$\uparrow$ (m)	$\updownarrow$ (-)	$\downarrow$ (%)	$\downarrow$ (%)	$\downarrow$ (%)
$[$ -1 , -1 , -1 , -1 $]$ , sing.-ax. train, joint guid.	100/ 0 0/ 0 0	2.28	+ 0.026	1.95	2.77	20.18
group + 1, joint train, joint guid.	100/ 0 0/ 0 0	2.43	+ 0.078	0.50	1.34	0 5.46
group + 1, joint train, per-ax. guid.	0 95/ 0 5/ 0 0	2.33	+ 0.193	0.40	2.96	19.35
yield + 1, $\sum_{i}w_{i}=1$	0 97/ 0 3/ 0 0	2.42	+ 0.046	0.29	1.00	14.64
yield + 1, $\sum_{i}w_{i}=3$	0 96/ 0 4/ 0 0	2.42	- 0.001	0.23	0.66	11.72

TABLE V : Ablation results for composition scheme, guidance weight, online expert policy and feasibility projection layer on a set of 500 geometric scenarios.

Fig. 6 : Physical demonstration results for the distinct left-side passing (left) and right-side passing (right) styles on a repeated test.

Robot Navigation

Robot navigation studies how robots perceive, localize, plan, and move safely toward goals in static or changing environments.

See all 554 papers

Performance-Preserving Online Adaptation in Social Navigation via Diffusion Steering
Sep 21, 2026Haruto Nagahisa, Kohei Matsumoto, Yuki Hyodo +1Robot Policy Adaptation Reinforcement Learning
Graduate School of Information Science and Electrical Engineering, Kyushu University, Fukuoka, Japan · Faculty of Information Science and Electrical Engineering, Kyushu University, Fukuoka, Japan
SAGE: A Socially-Aware Generative Engine for Heterogeneous Multi-Agent Navigation
Jul 18, 2026Lan Hu, Minghui Liwang, Wenbo Zhu +5Multi-Robot Motion Planning Multi-Agent Trajectory Prediction
Guohao School, Tongji University, Shanghai, China · Shanghai Research Institute for Intelligent Autonomous Systems, the State Key Laboratory of Autonomous Intelligent Unmanned Systems, Department of Control Science and Engineering, Tongji University, Shanghai, China · Department of Electrical and Computer Engineering, Western University, Ontario, Canada +2
E-SocialNav: Efficient Socially Compliant Navigation with Language Models
Mar 21, 2026Ling Xiao, Daeun Song, Xuesu Xiao +1Social Robot Navigation
Hokkaido University, Sapporo, Japan · George Mason University, Virginia, United States · The University of Tokyo, Tokyo, Japan

Sep 21, 2026J/K move · Enter open · S save

Performance-Preserving Online Adaptation in Social Navigation via Diffusion Steering

Haruto Nagahisa, Kohei Matsumoto, Yuki Hyodo, Ryo Kurazume

Graduate School of Information Science and Electrical Engineering, Kyushu University, Fukuoka, Japan · Faculty of Information Science and Electrical Engineering, Kyushu University, Fukuoka, Japan

Robot Policy Adaptation Reinforcement Learning Robotic RL Diffusion Policy

Open paper AskPDF arXiv

In social navigation, modeling the complex interactions between humans and robots is difficult, and deep reinforcement learning has therefore been actively studied. However, because simulation alone cannot fully reproduce diverse scenarios, robot dynamics, and the social conventions that vary across deployment environments, fine-tuning in the deployment environment is promising. In doing so, learning that preserves the base model's performance is required, so as not to compromise the primary objective of navigation, namely avoiding pedestrians and reaching the destination. In this study, we propose a method that applies diffusion steering via reinforcement learning (DSRL), which trains only the noise policy while keeping the diffusion policy fixed, thereby achieving learning that preserves performance. Furthermore, we integrate diffusion-based RL policies trained with multiple seeds to construct the base policy, improving learning performance. Our evaluation shows that, compared with other methods, the proposed method enables efficient learning while preserving performance, and we confirm flexible behavior control through adaptation to social conventions, as well as its effectiveness on a physical robot through hardware-in-the-loop simulation.

Abstract

Robot Navigation

Explore similar work

Performance-Preserving Online Adaptation in Social Navigation via Diffusion Steering

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

E-SocialNav: Efficient Socially Compliant Navigation with Language Models