Embodied Domain Adaptation for Object Detection

Mobile robots rely on object detectors for perception and object localization in indoor environments. However, standard closed-set methods struggle to handle the diverse objects and dynamic conditions encountered in real homes and labs. Open-vocabulary object detection (OVOD), driven by Vision Language Models (VLMs), extends beyond fixed labels but still struggles with domain shifts in indoor environments. We introduce a Source-Free Domain Adaptation (SFDA) approach that adapts a pre-trained model without accessing source data. We refine pseudo labels via temporal clustering, employ multi-scale threshold fusion, and apply a Mean Teacher framework with contrastive learning. Our Embodied Domain Adaptation for Object Detection (EDAOD) benchmark evaluates adaptation under sequential changes in lighting, layout, and object diversity. Our experiments show significant gains in zero-shot detection performance and flexible adaptation to dynamic indoor conditions.

Authors

• Xiangyu Shi
• Yanyuan Qiao
• Lingqiao Liu
• Feras Dayoub

Keywords

• Location awareness
• Lighting
• Object detection
• Contrastive learning
• Benchmark testing
• Performance gain
• Robot learning
• Indoor environment
• Mobile robots
• Standards
• Domain Adaptation
• Domain Shift
• Indoor Environments
• Mobile Robot
• Self-supervised Learning
• Temporal Clustering
• Pseudo Labels
• Positive Samples
• Negative Samples
• Adaptive Method
• Dynamic Environment
• Teacher Model
• Diverse Environments
• Simulation Environment
• Bounding Box
• Kullback-Leibler
• Target Domain
• Incremental Learning
• Challenging Conditions
• Source Domain
• Student Model
• Perception Of The Robot
• Merging Clusters
• Contrastive Loss
• Object Detection Dataset
• Source Model
• Temporal Consistency
• Consecutive Frames
• Region Proposal Network