CeDR-Nav: Cross-Embodiment Mapless Navigation via Offline Demonstration Recasting

Xianghui Wang; Xinming Zhang; Yiwei Qian; Siwei Cheng; Shanze Wang; Dan Zhang; Wei Zhang

cs.ROOct 9, 2026

CeDR-Nav: Cross-Embodiment Mapless Navigation via Offline Demonstration Recasting

Authors: Xianghui Wang, Xinming Zhang, Yiwei Qian, Siwei Cheng, Shanze Wang, Dan Zhang, Wei Zhang

Organizations: Department of Mechanical Engineering, The Hong Kong Polytechnic University, Hung Hom, Kowloon, Hong Kong · College of Information Science and Technology, Eastern Institute of Technology, Ningbo, China · School of Computer Science and Technology, University of Science and Technology of China, Hefei, China · Department of Mechanical Engineering, National University of Singapore, Singapore · Department of Aeronautical and Aviation Engineering, The Hong Kong Polytechnic University, Hung Hom, Kowloon, Hong Kong

Abstract

Learning-based methods have demonstrated strong performance in mapless robot navigation. However, a learned policy remains restricted to the embodiment on which it was trained. Many cross-embodiment navigation methods require time-consuming online training and can still underperform target-specific policies. In fact, demonstrations collected on a source embodiment already contain navigation experience that is potentially reusable for deriving a policy for each target embodiment, but the embodiment mismatch prevents these demonstrations from directly serving as valid supervision for the target. In this article, we present CeDR-Nav, a cross-embodiment demonstration recasting framework that trains a dedicated navigation policy for each target embodiment from existing source demonstrations without online data collection. Geometry-consistent scaling with per-demonstration validity assessment converts the source demonstrations into target training data, and a target-specific Gaussian mixture policy generates feasible avoidance maneuvers for each target embodiment. Extensive quantitative simulations and real-world deployments across multiple robot platforms demonstrate the effectiveness of CeDR-Nav, with an average success-rate gain of 13.8 percentage points over DRL-DCLP in cross-embodiment simulations and target-specific policy training completed within seconds.

Figures & tables

Fig. 1 : Overview of the CeDR-Nav framework.

Panel of Figure 2 — Fig. 2 : Validity assessment of a recast demonstration.

Panel of Figure 3 — Fig. 3 : Simulation scenarios used in the experiments.

$L_{1}\!=\!L_{2}$ (m)	$W_{1}\!=\!W_{2}$ (m)	SEnv1							SEnv2
		Map size	Ours			DRL-DCLP			Map size	Ours			DRL-DCLP
		Map size	SR	CR	TO	SR	CR	TO	Map size	SR	CR	TO	SR	CR	TO
0.300	0.300	$15{\times}15\,\text{m}^{2}$	94%	0%	6%	93%	0%	7%	$10{\times}10\,\text{m}^{2}$	100%	0%	0%	85%	1%	14%
0.600	0.300	$24{\times}24\,\text{m}^{2}$	93%	1%	6%	87%	0%	13%	$16{\times}16\,\text{m}^{2}$	95%	0%	5%	74%	0%	26%
0.600	0.450	$27{\times}27\,\text{m}^{2}$	91%	0%	9%	87%	0%	13%	$18{\times}18\,\text{m}^{2}$	95%	1%	4%	68%	0%	32%
0.700	0.300	$27{\times}27\,\text{m}^{2}$	96%	0%	4%	91%	0%	9%	$18{\times}18\,\text{m}^{2}$	100%	0%	0%	71%	0%	29%

TABLE I: Cross-embodiment navigation performance in simulation. SR, CR, and TO denote success rate, collision rate, and timeout rate, respectively. Results are averaged across three training seeds, each evaluated on 100 tasks per scenario.

Method	SR $\uparrow$	CR $\downarrow$	TO $\downarrow$	Metric $\uparrow$
DWA [ 28 ]	53.00%	6.00%	41.00%	0.2650
E-Band [ 29 ]	55.00%	12.67%	32.33%	0.2750
E2E [ 27 ]	50.67%	49.00%	0.33%	0.2533
DRL-DCLP [ 7 ]	84.00%	6.67%	9.33%	0.4200
CeDR-Nav (Ours)	86.33%	3.00%	10.67%	0.4313

TABLE II: Navigation performance of different methods on the BARN benchmark.

Panel of Figure 6 — Fig. 4 : Trajectory comparison in a representative BARN environment.

Fig. 5 : Wall-clock training time comparison between CeDR-Nav and DRL-DCLP across different GPU configurations.

$(L_{1}+L_{2})\times$ $(W_{1}+W_{2})$ (m)	w/o recasting	Fixed $\lambda_{\mathrm{ref}}$	Fixed $\lambda=1.5$	Plain MLP	Deploy- Scaling	CeDR-Nav
$0.6\times 0.6$	58.00	93.33	96.00	93.33	91.67	100.00
$1.2\times 0.6$	61.00	90.33	97.00	94.33	87.33	95.00
$1.2\times 0.9$	29.00	81.33	96.00	93.67	52.67	95.00
$1.4\times 0.6$	61.33	86.67	96.00	94.67	70.00	100.00
$1.4\times 0.9$	30.00	69.33	96.00	94.33	43.67	94.00
Average	47.87	84.20	96.20	94.07	69.07	96.80

TABLE III: Ablation studies and comparison with deployment-time scaling in SEnv2.

Fig. 6 : Trajectories of the tested robot configurations in the real-world experiments. The dimensions of the tested configurations are listed in Table IV . (a)–(d) TB2-C1 through TB2-C4 with CeDR-Nav. (e)–(h) TB2-C1 through TB2-C4 with DRL-DCLP under the same obstacle layouts. (i)–(k) Go2-C1 and Go2-C2 with CeDR-Nav, where (k) shows Go2-C2 during the dynamic test. (l) K1 with CeDR-Nav. The experimental videos can be found in the supplementary material.

	$L_{1}$ (m)	$L_{2}$ (m)	$W_{1}$ (m)	$W_{2}$ (m)	Sensor Pos. (m)
TB2-C1	0.200	0.200	0.700	0.300	0
TB2-C2	0.275	0.175	0.350	0.350	0
TB2-C3	0.300	0.700	0.200	0.200	0
TB2-C4	0.700	0.300	0.200	0.200	0
Go2-C1	0.350	0.350	0.155	0.155	0.187
Go2-C2	0.350	0.350	0.260	0.260	0.187

TABLE IV : Dimensions of the Tested Robot Configurations.

Config	Method	G1	G2	Path (m)
TB2-C1	CeDR-Nav	10/0/0	10/0/0	$3.56\pm 0.11$
	DRL-DCLP	10/0/0	0/9/1	$5.04\pm 2.56$
TB2-C2	CeDR-Nav	10/0/0	10/0/0	$5.38\pm 0.46$
	DRL-DCLP	0/10/0	—	$1.99\pm 0.15$
TB2-C3	CeDR-Nav	10/0/0	10/0/0	$3.94\pm 0.55$
	DRL-DCLP	10/0/0	0/10/0	$3.96\pm 0.13$

TABLE V: TurtleBot2 results. G1/G2: success/collision/timeout counts per 10 trials. Path: length to termination, including failures (mean

\pm

SD). “—”: G2 not entered after G1 failure.

Config	S/C/TO	Path (m)	Time (s)
Go2-C1	10/0/0	3.92 $\pm$ 0.69	16.19 $\pm$ 4.34
Go2-C2	10/0/0	5.48 $\pm$ 0.14	15.31 $\pm$ 1.29
Go2-C2 (Dyn)	10/0/0	5.62 $\pm$ 0.14	15.28 $\pm$ 1.10
K1	10/0/0	5.10 $\pm$ 0.32	14.39 $\pm$ 1.31

TABLE VI : CeDR-Nav results on Go2 and K1. S/C/TO are counts per 10 trials; path and time are mean

\pm

SD. Dyn denotes trials with a pedestrian entering the path.

Explore similar work

EA-Nav: Learning Safe Visual Navigation Policies with Embodiment Awareness
Jul 22, 2026Jialu Zhang, Yong Du, Xianda Guo +6Embodied Navigation Cross-Embodiment Robot Learning
Zhejiang University · Hangzhou, Zhejiang, China · Wuhan University +4
EmbodiSteer: Steering Embodiment-Agnostic Visuomotor Policies with Joint-Space Guidance for Zero-Shot Cross-Embodiment Deployment
Jun 11, 2026Shihefeng Wang, Kangchen Lv, Mingrui Yu +1Robot Skill Learning Generalization in Robotic Manipulation
Department of Automation, Tsinghua University · Beijing Key Laboratory of Embodied Intelligence Systems · Institute for Embodied Intelligence and Robotics, Tsinghua University
Cross-Embodiment Offline Reinforcement Learning for Heterogeneous Robot Datasets
Feb 20, 2026Haruki Abe, Takayuki Osa, Yusuke Mukuta +1RL for Robotics Cross-Embodiment Robot Learning
The University of Tokyo, Tokyo, Japan · RIKEN Center for Advanced Intelligence Project, Tokyo, Japan

Jul 22, 2026J/K move · Enter open · S save

EA-Nav: Learning Safe Visual Navigation Policies with Embodiment Awareness

Jialu Zhang, Yong Du, Xianda Guo, Shunwang Sun+5

Zhejiang University · Hangzhou, Zhejiang, China · Wuhan University+4

Embodied Navigation Cross-Embodiment Robot Learning Safe Robot Navigation Robot Imitation Learning

Open paper AskPDF arXiv

Cross-embodiment navigation is a key challenge in embodied intelligence. Due to differences in embodiment, the same visual observation may imply different actions for different agents, making prediction ambiguous when relying solely on vision. Existing studies mainly rely on reinforcement learning, which requires large-scale interaction and careful reward design, making it difficult to support scalable pretraining and real-world adaptation. In contrast, imitation-learning-based approaches remain limited. To address these challenges, we propose an imitation-learning-based embodiment-aware navigation framework with a modular multi-stage design. In pretraining, we construct a cross-embodiment navigation dataset from Internet videos and introduce embodiment geometry as conditional tokens to reduce action ambiguity under the same observation. In fine-tuning, we design a multimodal information injection mechanism based on a decoupled architecture. Specifically, we design a trajectory augmentation strategy to generate high-risk samples, which are used to train spatial perception and risk-aware correction separately, thereby explicitly incorporating embodiment geometry for safe navigation. Experimental results show that the proposed method effectively improves navigation performance across different embodiment settings, demonstrating the effectiveness of incorporating embodiment geometry into embodied navigation.

Abstract

Explore similar work

EA-Nav: Learning Safe Visual Navigation Policies with Embodiment Awareness

EmbodiSteer: Steering Embodiment-Agnostic Visuomotor Policies with Joint-Space Guidance for Zero-Shot Cross-Embodiment Deployment

Cross-Embodiment Offline Reinforcement Learning for Heterogeneous Robot Datasets