#Berkeley Robotics

Is It a Valuation Bubble or Real Demand? Investigating the Truth Behind Embodied Intelligence Data Collection Centers

The article explores the rapid expansion of embodied intelligence data collection centers in China, driven by both policy and market forces. It also reveals issues with some companies creating false revenue through a 'equipment sale—data repurchase' model. The article further analyzes the technical challenges of centralized data collection and the industry's shift toward distributed data collection models.

Towards a General Humanoid Loco-Manipulation Model via Egocentric Whole-Body Human Data Pretraining

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.

Whole-Body Aerial Grasping and Lifting via Partial Visual Observations

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.

Real-Time Whole-Body Safe Motion Generation for Multi-Segment Tendon-Driven Continuum Robots

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.

Bilinear World Models: Learning Representations with Structured Dynamics for Efficient Control

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.

KPI: A Promptable Kernel for Physical Interaction on Humanoids

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: Skill Assembly and Kinematic Imitation from Human Videos for Long-Horizon Mobile Manipulation

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.

Built Over 100 Data Training Sites, Why Are Robots Still 'Hungry'?

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.

The 'War' of Dexterous Hands: 8 Embodied Companies Show Off Their Skills

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.

Haozhi Qi@HaozhiQEditorial score6262

A new tool called Morphometric Imitation enables zero-shot sim-to-real visuomotor policies by retargeting human demonstrations to different multi-fingered robot hands, with a focus on morphology and contact awareness.

Sergey Levine@svlevineEditorial score2323

How do we run RL with real-time chunking (RTC)? In this work we figured out how to use a small RL policy with a large robot foundation model, where the RL policy observed more recent images (due to faster inference) and steers the policy toward better behaviors! A fun collaboration with Siemens, led by Brian Zhu, Momen Khalil, Emanuele Poggi from Siemens and @ehharrison4 from Berkeley, with lots of amazing contributors!

Robots Are Learning to Feel

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.