eNavi: Event-based Imitation Policies for Low-Light Indoor Mobile Robot Navigation

Prithvi Jai Ramesh; Kaustav Chanda; Krishna Vinod; Joseph Raj Vishal; Yezhou Yang; Bharatesh Chakravarthi

cs.ROMar 15, 2026

eNavi: Event-based Imitation Policies for Low-Light Indoor Mobile Robot Navigation

Authors: Prithvi Jai Ramesh, Kaustav Chanda, Krishna Vinod, Joseph Raj Vishal, Yezhou Yang, Bharatesh Chakravarthi

Organizations: School of Computing and Augmented Intelligence, Arizona State University, Tempe, AZ, USA.

Abstract

Event cameras provide high dynamic range and microsecond-level temporal resolution, making them well-suited for indoor robot navigation, where conventional RGB cameras degrade under fast motion or low-light conditions. Despite advances in event-based perception spanning detection, SLAM, and pose estimation, there remains limited research on end-to-end control policies that exploit the asynchronous nature of event streams. To address this gap, we introduce a real-world indoor person-following dataset collected using a TurtleBot 2 robot, featuring synchronized raw event streams, RGB frames, and expert control actions across multiple indoor maps, trajectories under both normal and low-light conditions. We further build a multimodal data preprocessing pipeline that temporally aligns event and RGB observations while reconstructing ground-truth actions from odometry to support high-quality imitation learning. Building on this dataset, we propose a late-fusion RGB-Event navigation policy that combines dual MobileNet encoders with a transformer-based fusion module trained via behavioral cloning. A systematic evaluation of RGB-only, Event-only, and RGB-Event fusion models across 12 training variations ranging from single-path imitation to general multi-path imitation shows that policies incorporating event data, particularly the fusion model, achieve improved robustness and lower action prediction error, especially in unseen low-light conditions where RGB-only models fail. We release the dataset, synchronization pipeline, and trained models at https://eventbasedvision.github.io/eNavi/

Figures & tables

Dataset	Environment	Sensor Platform	Primary Task	Contains Action ( $v$ , $\omega$ )?	Low-Light Focus?
MVSEC [ 1 ]	Indoor/Outdoor	Drone / Car	SLAM / Pose	×	×
DDD17 [ 2 ]	Outdoor	Car	Steering Prediction	✓	×
MMID [ 3 ]	Indoor	Humanoid	Depth Estimation	×	×
SEBVS [ 4 ]	Synthetic	Mobile Robot	Navigation	✓	×
eNavi (Ours)	Indoor	TurtleBot (Mobile)	Navigation	✓	✓

TABLE I: Summary of event-based datasets used in robotic perception and navigation tasks.

Fig. 1: Overview of the dataset generation workflow. (a) The process begins with teleoperated data collection using a mobile robot setup. (b) The collection process where the user is teleoperating the bot to follow another person across the room in three different paths ( $P_{1}$ , $P_{2}$ , and $P_{3}$ ) in different lighting conditions.

Fig. 2: Overview of the dataset generation workflow. The process begins with teleoperated data collection using a mobile robot setup. The recorded ROS $2$ bags are analyzed for trajectory and velocity consistency (center panels) before being synchronized into paired image, event, and action tuples $\{(\mathbf{I}_{k},\mathbf{E}_{k},\mathbf{a}_{k})\}$ stored in .h5 format.

Fig. 3: The Event-based Navigation Policy ( ENP-Fusion ) Architecture. The model fuses a frozen RGB stream and a trainable event stream using a Transformer-based attention mechanism. The system takes synchronized RGB frames ( $320\times 180\times 3$ ) and event frames ( $320\times 180\times 2$ ) as input and outputs continuous differential drive commands ( $v,\omega$ ) via an MLP head. The snowflake icon denotes that the RGB backbone weights remain fixed during training to prevent overfitting in low-light scenarios.

Variation	Modal	Train*	Best	MAE	Variation	Modal	Train*	Best	MAE
	RGB	50	48	0.0548		RGB	50	50	0.0667
	Event	37	29	0.0393		Event	44	36	0.0378
(Single-Path, Normal-Light)	Fusion	27	19	0.0388	(Multi-Path, Normal-Light)	Fusion	29	21	0.0402
	RGB	50	50	0.0584		RGB	50	50	0.0707
	Event	26	18	0.0388		Event	18	10	0.0416
(Single-Path, Mixed-Light)	Fusion	31	23	0.0358	(Multi-Path, Mixed-Light)	Fusion	30	22	0.0370

TABLE II: Validation metrics across all the training. All variations are trained for

50

epochs, where Train* indicates the model was run with early stopping at

8

epochs.

		Normal-Light			Low-Light
Method	Modal	Linear MAE	Angular MAE	Total MAE	Linear MAE	Angular MAE	Total MAE
Yolo8n + PID (Single-Path)	RGB	0.0882	0.0756	0.0819	0.0758	0.0667	0.0713
	RGB	0.0182	0.0237	0.0210	0.0258	0.0168	0.0213
	Event	0.0177	0.0275	0.0266	0.0234	0.0236	0.0235
Single-Path, Normal-Light (Ours)	Fusion	0.0183	0.0259	0.0221	0.0246	0.0199	0.0222
Yolo8n + PID (Multi-Path)	RGB	0.0841	0.1155	0.0998	0.0724	0.1342	0.1033

TABLE III: Evaluation of Normal-Light under varying illumination. The YOLOv8n + PID baseline is evaluated on its respective path type Single-Path (top) and Multi-Path (down) for direct comparison.

Fig. 4: Trajectory Analysis of ENP-Fusion Policy. Comparison of predicted linear velocity (m/s) and angular rate (rad/s) against expert ground truth for representative test trajectory ( $P_{3}$ ). (a) and (b) contrast performance under normal-light (top) and low-light (down) conditions.

Explore similar work

Event-VLA: Action-Conditioned Event Fusion for Robust Vision-Language-Action Model
Jun 28, 2026Jiaxin Liu, Xun Xu, Zhenhao Zhang +5Language-Conditioned Robot Manipulation Vision-Language-Action Models
ShanghaiTech · HKUST(GZ) · UMich +1
E-VLA: Event-Augmented Vision-Language-Action Model for Dark and Blurred Scenes
Apr 6, 2026Jiajun Zhai, Hao Shi, Shangwei Guo +2Event-Based Vision VLM Robustness
Zhejiang University · Ant Group · Hunan University
EVIS: Real-Time Event Camera Simulation with Multimodal Supervision in NVIDIA Isaac Sim
Jul 9, 2026Linli Shi, Ruijun Zhang, Ziyun WangEvent-Based Vision Robotics Simulation
Johns Hopkins University

Jun 28, 2026J/K move · Enter open · S save

Event-VLA: Action-Conditioned Event Fusion for Robust Vision-Language-Action Model

Jiaxin Liu, Xun Xu, Zhenhao Zhang, Hanqing Wang+4

ShanghaiTech · HKUST(GZ) · UMich+1

Language-Conditioned Robot Manipulation Vision-Language-Action Models Robustness of VLA Models Multimodal Sensor Fusion

Open paper AskPDF arXiv

Vision-Language-Action (VLA) models have become an important paradigm of embodied AI. However, existing VLA models typically assume well-lit and stable indoor settings, while real-world embodied manipulation may involve degraded RGB observations caused by illumination shifts, posing critical challenges for robust robotic manipulation. To address this gap, we propose Event-VLA, an event-enhanced VLA framework for generalizable manipulation across varying illumination conditions. We formulate VLA-based manipulation under degraded visibility as a practical robustness problem for RGB-centric policies, and introduce event streams as an illumination-robust, motion-sensitive complementary observation to improve robustness across visibility levels. Specifically, unlike conventional multimodal fusion that directly merges event features into the global semantic token space, Event-VLA injects event information through an action-query routing pathway. It uses learnable action queries to extract task-relevant semantics from the VLA reasoning process, and selectively aggregates event tokens via gated cross-attention to construct event-aware action representations. This design preserves the pretrained RGB-language semantic priors while effectively leveraging event information for robust action prediction. Experiments in simulation and real-world deployment show that Event-VLA maintains strong manipulation performance under normal lighting and improves success rates under low-light degradation and near-dark real-world settings.

Abstract

Explore similar work

Event-VLA: Action-Conditioned Event Fusion for Robust Vision-Language-Action Model

E-VLA: Event-Augmented Vision-Language-Action Model for Dark and Blurred Scenes

EVIS: Real-Time Event Camera Simulation with Multimodal Supervision in NVIDIA Isaac Sim