The paper presents HumanVerse-500, a 500-hour dataset of human loco-manipulation behaviors collected with a lightweight wearable system. It introduces λ₀, a whole-body humanoid vision-language-action policy trained through three stages, achieving state-of-the-art performance on real-world tasks and analyzing how human data supports downstream control.
This research presents a recurrent teacher-student framework that learns a single policy in simulation to jointly command flight, arm motion, and gripper closure without explicit task-phase input. The method achieves high success rates in simulation under nominal, physics/control-randomized, and additional camera-randomized conditions.
This research presents a unified actuation-space framework for planar multi-segment tendon-driven continuum robots, combining energy-based modeling with constraint satisfaction techniques to achieve real-time whole-body safe motion generation. The method achieves collision-free success rates of 96% and 100% in static and dynamic scenarios, respectively, compared to 60% and 80% without CBF constraints. The model closely matches GVS references with a maximum curvature error of $5.223 imes 10^{-2} ext{m}^{-1}$, and demonstrates real-time performance with a mean control-step time of 6.66 ms.
This research proposes a JEPA-style world model with bilinear parameterization of latent dynamics, enabling efficient planning and control. The structured approach allows for action recoverability and prevents representation collapse, demonstrating success in both long-horizon planning and real-time control tasks.
This research introduces KPI, a promptable kernel for physical interaction on humanoids, which allows trajectory sources to send contracts specifying force ranges and interaction behaviors. The kernel adapts stiffness, damping, and reference values in real-time based on tracking error and wrench estimates, enabling instruction-driven tasks like winch operation and door opening without task-specific code.
SAKI is a framework that enables robots to acquire and compose manipulation skills from human videos, allowing for long-horizon mobile tasks. It integrates cross-demonstration assembly, whole-body kinematic imitation, and persistent object estimation. The method demonstrates skill reuse across layouts and the composition of independent interactions into continuous tasks, with ablation studies showing the effectiveness of task-conditioned reference preparation.
The article explores the issue of data scarcity in the field of embodied intelligence, pointing out that although there are already a large number of data collection centers, high-quality and reusable data remains severely insufficient. The article analyzes structural contradictions such as high data collection costs, inconsistent quality, fragmented formats, and lack of cross-ontology reusability. It introduces Wu Wen Tech's solution of building a data foundation through a Real2Sim2Real closed-loop system, including large-scale data collection, automated annotation, and simulation training technologies, ultimately forming a data-driven flywheel to promote the development of embodied intelligence.
Editorial context：The article highlights the critical data scarcity in embodied AI, emphasizing the gap between the scale of robot models and the availability of high-quality physical interaction data. It identifies structural issues such as high collection costs, quality inconsistencies, and lack of standardization as major barriers. The solution proposed by Wu Wen Tech involves a Real2Sim2Real closed-loop system,
This article explores the complexity of dexterous hands in real-world tasks, pointing out that evaluation criteria have shifted from single parameters to overall operational capabilities. Eight companies analyze, from standards, hardware, perception, data, models to commercialization, how to achieve the reliability and practicality of dexterous hands. The article emphasizes the integration of vision and tactile sensing, the importance of data loops, and the challenges between model training and real-world deployment.
Editorial context：This article provides a comprehensive overview of the challenges and approaches in evaluating and developing dexterous robotic hands, emphasizing the integration of hardware, perception, data, and models. It highlights the need for industry standards, the role of vision and tactile sensing, and the importance of data collection and model training for real-world deployment.
Haozhi Qi@HaozhiQEditorial score6262One human demonstration. Any multi-fingered hand. Zero-shot sim-to-real visuomotor policy. https://morphometricimitation.github.io Collaborators: @he_siming @ckwolfeofficial @HaozhiQ @LeaMue27 Shankar Sastry, Claire Tomlin, @JitendraMalikCV
Sergey Levine@svlevineEditorial score2323
Asynchronous VLA inference reduces inference delay, but breaks the Markovian assumption necessary for RL fine-tuning. How can we enable RL fine-tuning of VLAs with async inference? We introduce ARLI: Asynchronous RL with Intermediate Information! https://async-rl-intermediate-information.github.io/ (1/n)
This article explores how tactile data is helping robots improve their dexterous manipulation skills. Researchers are creating large tactile datasets and models that can use tactile feedback to enhance robot performance in tasks like folding laundry or turning keys. Challenges include the difficulty of integrating tactile data with vision-based models and the need for more diverse and scalable datasets to achieve significant improvements in robot dexterity.
Editorial context：The article highlights the growing importance of tactile data in advancing robot dexterity, emphasizing the challenges of integrating tactile feedback into vision-language-action (VLA) models. It discusses recent research efforts to create diverse tactile datasets and models that can generalize across different robotic hardware, while also addressing the limitations of current approaches and the '