Humanoid robots are drawing significant attention as versatile platforms for complex motor control, human-robot interaction, and general-purpose physical intelligence. However, achieving efficient whole-body control (WBC) in humanoids remains a fundamental challenge due to sophisticated dynamics, underactuation, and diverse task requirements. While learning-based controllers have shown promise for complex tasks, their reliance on labor-intensive and costly retraining for new scenarios limits real-world applicability. To address these limitations, behavior(al) foundation models (BFMs) have emerged as a new paradigm that leverages large-scale pre-training to learn reusable primitive skills and broad behavioral priors, enabling zero-shot or rapid adaptation to a wide range of downstream tasks. In this paper, we present a comprehensive overview of BFMs for humanoid WBC, tracing their development across diverse pre-training pipelines. Furthermore, we discuss real-world applications, current limitations, urgent challenges, and future opportunities, positioning BFMs as a key approach toward scalable and general-purpose humanoid intelligence. Finally, we provide a curated and regularly updated collection of BFM papers and projects to facilitate further research, which is available at https://github.com/yuanmingqi/awesome-bfm-papers.

Continuum robots(CRs) are featured with high compliance and adaptability. However, existing CRs commonly face issues of insufficient stiffness and actuation redundancy. This paper proposes a novel design of multi-layer jamming spherical joint (MLJSJ) and develops a pneumatically actuated variable-stiffness continuum robot. A segmented spherical expansion actuator airbag, resembling Tanghulu, is fabricated via the lost-wax method to meet the positive-pressure actuation requirements. The proposed robot consists of four serially connected continuum segments, each containing four spherical joints for smooth bending motion. By leveraging multi-layer frictional interactions within jamming structures, the robot achieves rapid and wide-range stiffness modulation under relatively low air pressure. Each continuum segment can be locked via a dedicated actuation airbag, with its pressure regulated by a proportional valve to enable rapid switching between rigid and flexible states. By switching the lock-unlock states of segments, the variable-stiffness mechanism decouples multi-segment motions into single-segment bending, enabling sequential actuation of all segments by using only four embedded alloy rods. A dedicated experimental setup was developed to characterize performance, demonstrating that a single segment reaches a maximum stiffness of 1403.03 N/m, with a stiffness modulation ratio of 35.85 and switching times ranging from 0.07 s to 0.43 s. Furthermore, by controlling the stiffness states and bending motions of different segments, the single segment of robot demonstrates a maximum payload capacity of 2 kg and versatile motion flexibility verified through obstacle avoidance and grasping tasks.

Graph Neural Networks (GNNs) garnered significant success in modeling graphs due to their powerful capabilities in representation learning and reasoning. However, the lack of explainability in GNNs largely limits their application in security, sensitive, and other such scenarios. Due to their interpretability and high fidelity, surrogate methods in explainability techniques have sparked some explorations, but many challenges are still not well-addressed. 1) Most surrogate based explainers generate local neighborhoods by randomly perturbing node features, ignoring the graph topology perturbations. 2) As the graph grows, the combinations of perturbations increase exponentially, whereas only a few perturbation combinations are close to the prediction. 3) The application of multiple relaxation steps and regularization terms in practice may diminish the intrinsic interpretability of surrogate models in previous works. Towards this end, we propose a novel framework named MIP-Explainer, which consists of Selector and Explainer modules. In the Selector module, we leverage mutual information to perturb the graph structure, efficiently generating local neighborhoods of the data. In the Explainer module, we propose an intuitive and interpretable linear model to these local neighborhoods, without incorporating any relaxation or regularization terms. Extensive experiments on real-world and synthetic datasets show the effectiveness of our MIP-Explainer by outperforming state-of-the-art baselines.

Advances in robotics have been driving the development of human-robot interaction (HRI) technologies. However, accurately perceiving human actions and achieving adaptive control remains a challenge in facilitating seamless coordination between human and robotic movements. In this paper, we propose a hierarchical procedural framework to enable dynamic robot-assisted hand-object interaction (HOI). An open-loop hierarchy leverages the RGB-based 3D reconstruction of the human hand, based on which motion primitives have been designed to translate hand motions into robotic actions. The low-level coordination hierarchy fine-tunes the robot’s action by using the continuously updated 3D hand models. Experimental validation demonstrates the effectiveness of the hierarchical control architecture. The adaptive coordination between human and robot behavior has achieved a delay of ≤ 0.3 seconds in the tele-interaction scenario. A case study of ring-wearing tasks indicates the potential application of this work in assistive technologies such as healthcare and manufacturing.

Graph neural networks (GNNs) have shown significant advantages in learning structured data representations and have become one of the important methods in machine learning. However, the inherent opacity of deep learning models hinders their widespread application in critical domains, making the explanation of GNNs predictions particularly crucial. While explanations for supervised learning models have been extensively studied, research on generating explanations for unsupervised learning models remains relatively scarce. Current explanation methods for unsupervised models mostly focus on either factual reasoning or counterfactual reasoning. But relying on a single reasoning approach often results in explanations that are either redundant or incomplete. To address this challenge, this paper proposes a novel explanation method for Unsupervised Models named UM-Explainer, which is based on factual and counter-factual reasoning. By calculating the distance between cluster centers and the example to be explained, a contrastive loss based on both factual and counterfactual reasoning is constructed to optimize the mask, thereby generating an ideal explanatory subgraph. To validate the effectiveness of UM-Explainer, we conducted experiments with synthetic and real datasets on multiple unsupervised models. Experimental results demonstrate that UM-Explainer performs outstandingly in generating accurate explanations.

When evaluating a learner's knowledge proficiency, the multiple-choice question is an efficient and widely used format in standardized tests. Nevertheless, generating these questions, particularly plausible distractors (incorrect options), poses a considerable challenge. Generally, the distractor generation can be classified into cloze-style distractor generation (CDG) and natural questions distractor generation (NQDG). In contrast to the CDG, utilizing pre-trained language models (PLMs) for NQDG presents three primary challenges: (1) PLMs are typically trained to generate “correct” content, like answers, while rarely trained to generate “plausible” content, like distractors; (2) PLMs often struggle to produce content that aligns well with specific knowledge and the style of exams; (3) NQDG necessitates the model to produce longer, context-sensitive, and question-relevant distractors. In this study, we introduce a fine-tuning framework named DGRC for NQDG in Chinese multi-choice reading comprehension from authentic examinations. DGRC comprises three major components: hard chain-of-thought, multi-task learning, and generation mask patterns. The experiment results demonstrate that DGRC significantly enhances generation performance, achieving a more than 2.5-fold improvement in BLEU scores.

Recent advances in machine learning have paved the way for the development of musical and entertainment robots. However, human–robot cooperative instrument playing remains a challenge, particularly due to the intricate motor coordination and temporal synchronization. In this article, we propose a theoretical framework for human–robot cooperative piano playing based on nonverbal cues. First, we present a music improvisation model that employs a recurrent neural network (RNN) to predict appropriate chord progressions based on the human's melodic input. Second, we propose a behavior-adaptive controller to facilitate seamless temporal synchronization, allowing the cobot to generate harmonious acoustics. The collaboration takes into account the bidirectional information flow between the human and robot. We have developed an entropy-based system to assess the quality of cooperation by analyzing the impact of different communication modalities during human–robot collaboration. Experiments demonstrate that our RNN-based improvisation can achieve a 93% accuracy rate. Meanwhile, with the MPC adaptive controller, the robot could respond to the human teammate in homophony performances with real-time accompaniment. Our designed framework has been validated to be effective in allowing humans and robots to work collaboratively in the artistic piano-playing task.