
ISSN: 2960-1436 (Print)
ISSN: 2960-1444 (Online)
CODEN: RLABAV
For any inquiries regarding journal development, the peer review process, copyright matters, or other general questions, please contact the editorial office.
E-Mail: rl@elspub.com
For production or technical issues, please contact the production team.
E-Mail: production@elspub.com
Embodied intelligence extends artificial intelligence from information processing in the digital domain to autonomous interaction with the physical world. Despite rapid progress, the field still faces challenges such as ambiguous conceptual boundaries, fragmented technical frameworks, and insufficient alignment between application requirements and evaluation standards. This review presents a systematic framework for understanding embodied intelligence from five perspectives: conceptual foundations, key technologies, representative applications, evaluation methodologies, and future challenges. We first distinguish embodied intelligence from teleoperation and conventional automation, and establish an evolutionary framework spanning disembodied intelligence, spatial intelligence, physical intelligence, and embodied intelligence. Then we summarize key technologies in embodiment, the cerebellum, brain architectures, data acquisition, and training approaches, with particular emphasis on the evolution from traditional hierarchical modular architecture to large-model-driven modular architecture, Vision-Language-Action (VLA) model architecture, world model architecture, and world action model architecture. Representative applications are categorized into structured-environment manipulation, unstructured-environment interaction, and high-risk open-domain exploration. To bridge technological development and real-world deployment, we further review three evaluation paradigms which include data-based evaluation, simulation-based testing and real-world testing. Finally, we identify eight major challenges which cover data quality, long-horizon planning, multimodal fusion, edge deployment, safety and trustworthiness, continual learning, unified navigation and manipulation, and multi-agent collaboration. This review provides a systematic perspective on embodied intelligence and highlights key directions for advancing embodied systems toward reliable deployment in complex real-world environments.
Egocentric interaction grounding requires a system to identify one target when several alternatives are plausible. We propose a framework that combines video-conditioned interaction hypotheses, localization-guided direct preference optimization (DPO), confidence-weighted box fusion, and exponential moving average (EMA) smoothing. We construct preference pairs from training-set localization quality to adapt the hypothesis generator, then combine grounded hypotheses using detector confidence at inference. On RefEgo, our method improves final-output mean intersection over union (mIoU) from 0.177 to 0.232 compared with an adapted reasoning-and-grounding baseline. The detector-proposal oracle reference increases from 0.422 to 0.434, while the reference-to-output gap decreases from 0.245 to 0.202. The fractions of target-present evaluation records meeting intersection over union thresholds of 0.5 and 0.7 are 0.257 and 0.193, respectively. Candidate-generation analysis, a separate pre-/post-DPO comparison, and component ablations complement the main comparison. These results support combining localization-guided preference learning with confidence-weighted spatial aggregation for more accurate final target localization in egocentric video.
Skin lesion segmentation supports computer-aided dermatology, but blurred boundaries, low contrast, and marked variation in dermoscopic appearance make reliable delineation difficult. Foundation models such as the Segment Anything Model (SAM) offer strong visual priors, although their sensitivity to prompt quality limits their reliability in few-shot medical settings. We present PADS, a novel perturbation-aware distillation framework for robust skin lesion segmentation under prompt variability. During training, we perturb bounding-box prompts derived from ground-truth masks to simulate realistic prompt noise. A lightweight relevance module then estimates channel-wise feature importance under these perturbations and guides selective distillation from MedSAM into a compact UNet. This training improves robustness to imperfect or noisy prompts, allowing pseudo-prompts to support reliable segmentation. At inference, the UNet either produces the segmentation directly or supplies a pseudo-prompt for optional MedSAM refinement. Experiments on ISIC 2018 and cross-dataset evaluation on PH2 show improved few-shot segmentation accuracy and slower performance degradation as prompt perturbation increases, compared with standard distillation. PADS adds only limited computational overhead relative to MedSAM. The experimental results support perturbation-aware distillation as a practical approach to robust few-shot medical image segmentation.