FLUX: Accelerating Cross-Embodiment Generative Navigation Policies via Rectified Flow and Static-to-Dynamic Learning

Zeying Gong; Yangyi Zhong; Yiyi Ding; Tianshuai Hu; Guoyang Zhao; Lingdong Kong; Rong Li; Jiadi You

cs.ROMar 13, 2026

FLUX: Accelerating Cross-Embodiment Generative Navigation Policies via Rectified Flow and Static-to-Dynamic Learning

Authors: Zeying Gong, Yangyi Zhong, Yiyi Ding, Tianshuai Hu, Guoyang Zhao, Lingdong Kong, Rong Li, Jiadi You,

Organizations: The Hong Kong University of Science and Technology (Guangzhou) · The Hong Kong University of Science and Technology · National University of Singapore

Abstract

Autonomous navigation requires a broad spectrum of skills, from static goal-reaching to dynamic social traversal, yet evaluation remains fragmented across disparate protocols. We introduce DynBench, a dynamic navigation benchmark featuring physically valid crowd simulation. Combined with existing static protocols, it supports comprehensive evaluation across six fundamental navigation tasks. Within this framework, we propose FLUX, the first flow-based unified navigation policy. By linearizing probability flow, FLUX replaces iterative denoising with straight-line trajectories, improving per-step inference efficiency by 47% over prior flow-based methods and 29% over diffusion-based ones. Following a static-to-dynamic curriculum, FLUX initially establishes geometric priors and is subsequently refined through reinforcement learning in dynamic social environments. This regime not only strengthens socially-aware navigation but also enhances static task robustness by capturing recovery behaviors through stochastic action distributions. FLUX achieves state-of-the-art performance across all tasks and demonstrates zero-shot sim-to-real transfer on wheeled, quadrupedal, and humanoid platforms without any fine-tuning.

Figures & tables

Fig. 1 : FLUX : A Flow-Based Unified Policy for Cross-Embodiment Navigation. Our static-to-dynamic training curriculum enables efficient, socially-aware navigation, which transfers zero-shot across three heterogeneous platforms in the real world.

Fig. 2 : Overview of FLUX framework. FLUX follows a static-to-dynamic training paradigm. Stage 1 (Top): Given egocentric visual observations and a goal, the flow policy head is pre-trained via imitation learning on static expert trajectories. It generates diverse candidate paths, which are evaluated by the critic head. Stage 2 (Bottom): The framework is post-training using Group Relative Policy Optimization via on-policy rollouts. This stage optimizes for both goal-reaching efficiency and social compliance in dynamic environments.

Fig. 3 : Overview of DynBench evaluation suite. Our benchmark provides a unified framework for fundamental navigation. (a) Diverse environments spanning unstructured cluttered environments with randomized obstacles and structured Isaac Sim scenes. (b) Six task modalities categorized into goal-conditioned navigation and coverage maximization, supporting our static-to-dynamic learning paradigm.

Benchmark	Rendering	Physics	Behaviors	Planners
CrowdNav [ 17 ]	2D	$\times$	Rule-based	ORCA [ 22 ]
Arena-Bench [ 18 ]	2D&3D	$\times$	Random Walk	ORCA [ 22 ]
Social-HM3D [ 2 ]	3D	$\times$	Rule-based	A* [ 24 ] + ORCA [ 22 ]
URBAN-SIM [ 25 ]	3D	$\checkmark$	Rule-based	A* [ 24 ] + ORCA [ 22 ]
DynBench (Ours)	3D	$\checkmark$	Markov FSM	NavMesh-A* + Geometric [ 23 ]

TABLE I : Comparison of simulation benchmarks with dynamic humans. DynBench combines hierarchical Markov behavior modeling with NavMesh-based planning and geometry-based local avoidance, producing more naturalistic pedestrian interactions.

Methods	Static Scenes						Dynamic Scenes
	PointNav		Exploration		ImageNav		Dyn. PointNav		Dyn. Exploration		SocialNav
	SR $\uparrow$	SPL $\uparrow$	ET $\uparrow$	EA $\uparrow$	SR $\uparrow$	SPL $\uparrow$	SR $\uparrow$	S-TL $\downarrow$	ET $\uparrow$	EA $\uparrow$	SR $\uparrow$	Coll. $\downarrow$	SC $\downarrow$	MinDist. $\uparrow$
Traditional Planning Methods
$\bullet$ DWA [ 9 ]	33.5	32.4	-	-	-	-	38.2	18.5	-	-	41.8	10	7.3	4.1
$\bullet$ FBE [ 10 ]	-	-	4.3	11.2	-	-	-	-	20.3	39.5	-	-	-	-
Reinforcement Learning Methods

TABLE II : Performance comparison across six fundamental navigation tasks. Our method achieves state-of-the-art performance.

Fig. 4 : Real-World Cross-Embodiment Deployment. FLUX is deployed on three heterogeneous platforms—wheeled (Go2-W), quadrupedal (Go2), and humanoid (G1)—using only a single RGB-D camera. Despite significant differences in morphology and locomotion, our unified policy achieves robust zero-shot transfer without platform-specific fine-tuning. Critic-guided visualizations further show that safer paths with larger clearance from obstacles receive higher scores, demonstrating reliable risk-aware navigation in real-world scenarios.

Method	Head Type	Training	PointNav		SocialNav				Exploration		Dyn. Exploration
Method	Head Type	Training	SR $\uparrow$	SPL $\uparrow$	SR $\uparrow$	Coll. $\downarrow$	SC $\downarrow$	MinDist. $\uparrow$	ET $\uparrow$	EA $\uparrow$	ET $\uparrow$	EA $\uparrow$
$-$ Determinisitic Baseline ‡	Regression	IL	53.6	50.0	49.3	18	4.9	3.8	32.8	48.1	30.5	71.0
DDPM Baseline	Diffusion	IL	77.8	74.8	59.3	21	5.5	3.5	72.5	167.2	43.6	93.7
$+$ Conditional Flow Matching	Flow	IL	78.9	75.6	60.6	20	5.9	3.3	71.8	166.8	43.8	95.2
$+$ Rectified Flow	Flow	IL	80.4	78.0	62.8	17	5.8	3.6	72.8	168.5	48.0	106.4
$+$ GRPO Post-Training	Flow	IL $+$ RL	80.9	78.6	64.0	16	4.4	4.0	73.7	172.8	55.1	128.7

TABLE III : Ablation studies on FLUX . We report metrics for tasks most relevant to each component: PointNav and Exploration evaluate IL fine-tuning, while SocialNav and Dyn. Exploration assess RL post-training. ‡ indicates excluded setups.

Method	Type	Steps $K$	Inference Time (ms)
NoMaD [ 5 ]	Diffusion-based	10	468.4
NavDP [ 7 ]	Diffusion-based	10	305.2
FlowNav [ 6 ]	Flow-based	10	408.6
FLUX †		10	244.9
FLUX		6	216.5

TABLE IV : Inference efficiency comparison of generative navigation policies. FLUX † uses 10 Euler steps for fair comparison.

Fig. 5 : Sensitivity analysis of trajectory number $M$ and sample steps $K$ . FLUX outperforms NavDP across nearly all $(M,K)$ configurations, with $M{=}16$ , $K{=}6$ achieving near-peak performance.

Fig. 6 : Qualitative analysis of generative navigation policies. Compared to prior methods, FLUX generates trajectories with higher safety in both static and dynamic obstacle avoidance.

Robot Navigation Benchmarks

See all 26 papers

Learning Social Navigation from Internet Videos in the Policy State Space
Sep 26, 2026Jiaming Wang, Duc Thang Nguyen, Jizhuo Chen +4Robot Policy Learning Robot Navigation
National University of Singapore, Singapore · Singapore Management University, Singapore · The Chinese University of Hong Kong, Hong Kong SAR, China +1
Beyond Isolation: A Unified Benchmark for General-Purpose Navigation
May 10, 2026Samson Sun, Tianyi Yang, Tengyue Wang +6Embodied Navigation Cross-Embodiment Robot Learning
1AutoLab, SAI, Shanghai Jiao Tong University · 2Research Lab, Anyverse Dynamics
KinematicRL: A Sim-to-Real Reinforcement Learning Framework For Social Navigation With Kinodynamic Feasibility
Jun 10, 2026Zhiming Xu, Haodong Yang, Chengju Liu +2Kinodynamic Motion Planning Sim-to-Real Transfer
School of Computer Science and Technology, Tongji University, Shanghai 201804, China · Department of Electronics and Information Engineering, Tongji University, Shanghai 201804, China · Shanghai Institute of Intelligent Science and Technology, Tongji University, Shanghai 201210, China

Sep 26, 2026J/K move · Enter open · S save

Learning Social Navigation from Internet Videos in the Policy State Space

Jiaming Wang, Duc Thang Nguyen, Jizhuo Chen, Volodymyr Shcherbyna+3

National University of Singapore, Singapore · Singapore Management University, Singapore · The Chinese University of Hong Kong, Hong Kong SAR, China+1

Robot Policy Learning Robot Navigation Social Robot Navigation Robot Learning Robot Learning from Video

Open paper AskPDF arXiv

Training robust social-navigation policies requires simulators with diverse scene layouts, terrain, and human motion, but constructing such environments and specifying pedestrian behavior is costly. We propose an efficient pipeline that converts ordinary monocular walking videos directly into closed-loop social-navigation training environments in the policy's state space. Our key observation is that local social navigation primarily depends on two types of information: where the robot can traverse and how nearby pedestrians move. We therefore represent the static scene as a metric traversability map, which can be rigidly transformed under counterfactual robot motion, while directly replaying the pedestrian trajectories recovered from the video over time. This abstraction allows us to define the forward dynamics directly in the policy's state space and efficiently simulate counterfactual robot states without reconstructing or rendering photorealistic observations. The resulting policy achieves 81.2% success in the independent Arena benchmark, compared with 75.0% for the strongest baseline, and succeeds in 19/20 real-robot trials without policy fine-tuning. Project page: https://jiaming.im/VideoSocNav

Abstract

Robot Navigation Benchmarks

Explore similar work

Learning Social Navigation from Internet Videos in the Policy State Space

Beyond Isolation: A Unified Benchmark for General-Purpose Navigation

KinematicRL: A Sim-to-Real Reinforcement Learning Framework For Social Navigation With Kinodynamic Feasibility