Technical Directions
Robotics coverage relating to reinforcement learning.
89 selected reportsLast 30 days 78 recordsAll coverage 577 records
TouchScale is a 500-hour human vision-touch dataset collected with unified sensors and protocols, enabling scaling benefits in perception and robotic manipulation. The dataset improves cross-sensor generalization and robot task success rates without using action labels, demonstrating the potential of large-scale human touch data for embodied AI.
TouchScale is a 500-hour human visual-tactile dataset collected with unified sensors and synchronization, showing scaling benefits in both perception and robotic manipulation. The dataset enables cross-sensor generalization and improves robot task success rates from 22.5% to 57.5% through mid-training. The study highlights the potential of large-scale visual-tactile data to enhance embodied AI systems.
The article explores the rise of embodied intelligence as a leading AI field in 2025, explaining its technical stack, open-source ecosystems, and learning paths. It discusses the integration of perception, decision-making, and control in robotics, the role of open-source projects in reducing entry barriers, and the importance of data sharing and simulation-to-real transitions.
Researchers from ETH Zurich have developed a self-contained robotic hand that can walk on its fingertips and manipulate objects, using reinforcement learning and off-the-shelf hardware. The hand can navigate different surfaces, right itself, and push objects autonomously, suggesting new possibilities for modular and autonomous robotic systems.
Wuhan Huilike Robotics Co., Ltd. has developed the VLI multimodal embodied interaction model based on research achievements from a team at Huazhong University of Science and Technology. The company focuses on enhancing robots' social capabilities in real-world scenarios. Through an L0 to L5 interaction capability grading framework, the company defines multiple levels of social capabilities ranging from tool execution to non-cooperative strong active interaction. In specialized scenarios such as autism assistance intervention, Huilike combines expert experience with multimodal data to enable robots to achieve the goals of understanding humans, adjusting strategies, and providing continuous service.
This article explains the 'hand-eye-brain' coordination framework for industrial and service robots, detailing the roles of perception, decision-making, and execution. It covers challenges like sensor calibration, force control, and system synchronization, and offers practical advice on simulation, testing, and deployment in real-world scenarios.
This tutorial details the progression from imitation learning to Vision-Language-Action (VLA) models in robotics, focusing on the shift from repeating motions to understanding and executing tasks. It outlines key experiments with models like SmolVLA and π0.5, and discusses challenges such as data quality, model capacity, and simulation-to-real transfer, ultimately aiming to build robots that can see, understand, adapt, and act.
This article details the progression of a robotics team's work from simple imitation learning to developing Vision-Language-Action (VLA) models. It outlines challenges in adapting robotic behavior to new environments, the development of models like SmolVLA and π0.5, and the shift toward simulation-generated data for training. The focus is on building robots that can see, understand, and act autonomously in dynamic settings.
This paper presents a collection of recent research papers and datasets focused on vision-language-action (VLA) models in robotics. Topics include recursive video in-context learning, simulation-guided world models, closed-loop benchmarks for driving, and methods for improving generalization and safety in VLA systems. The work highlights challenges such as task success vs. instruction following, the grounding gap, and the need for efficient and reliable adaptation in real-world scenarios.
This guide outlines the algorithm stack for embodied AI, from engineering tools and perception to learning and decision-making. It emphasizes the importance of system reliability and data quality, and provides a structured learning path for beginners, including key resources and practical advice.
At the 2026 World Robotics Conference, Sergey Levine pointed out that the robotics industry is undergoing a shift from controlling the body to developing decision-making capabilities. He emphasized that robots need to combine prior knowledge with real-world experience to achieve true intelligence. Although robots are still in the foundational technology stage, through data flywheels and scaling, they may eventually enable broader applications.
This paper presents a comprehensive review of diffusion models in robotics, focusing on their application in robot learning, data scaling, reinforcement learning, and imitation learning. The authors conducted a systematic literature review, selecting peer-reviewed articles and high-impact preprints to analyze the current state of diffusion model research in robotics. The paper discusses various applications of diffusion models, including semantic-level data augmentation, cross-view and morphology synthesis, automated simulation asset and task generation, trajectory optimization, policy representation, hierarchical planning, and real-time deployment. It also addresses the challenges and limitations of current approaches, such as computational cost, inference latency, and data efficiency, while highlighting promising future directions for research and development in the field.
This paper presents LeaP, a learnable source prior for generative robot policies, which replaces fixed standard Gaussian distributions with state-dependent random initialization. LeaP jointly learns the mean and variance of an initial Gaussian distribution, improving action generation in robotic tasks. The method achieves an average success rate of 81.6% in simulation and 80.0% on real robotic arms, outperforming existing baselines like A2A and VITA. The paper is accepted at CoRL 2026, and the code is open-sourced.
Professor Shi Qing's team from Beijing Institute of Technology was inspired by the high-speed running mechanism of the elephant shrew. They designed a micro quadruped robot named FLEXOR equipped with a dual-joint coupled spine. Through dynamic spine-leg coordination, the robot achieved a 31.6% increase in speed and a 32.2% reduction in energy consumption without upgrading its motors. The study reveals the critical role of the timing coordination between the spine and legs in locomotion performance and verifies the universality of this mechanism across different robot morphologies, providing new insights for highly mobile embodied intelligent robots.
Aaron Ames, a leading roboticist at Caltech, argues that safety must be the primary concern for humanoid robots. He emphasizes the need for formal mathematical guarantees, such as control barrier functions (CBFs), to ensure safe operation. Ames highlights the gap between impressive demonstrations and real-world safety, urging the industry to prioritize safety standards to prevent harm.
The Stanford team's HomeBody project demonstrates how GPT-6 Astra controls the Unitree G1 robot to complete kitchen cleaning tasks through a three-step process: spatial exploration, digital twin simulation, and task execution. The system uses pre-trained models and a skill library to perform complex operations in new environments without additional training.
The article analyzes China's leading position in the field of humanoid robot patents, noting that as of the first half of 2026, China holds 60.8% of relevant patents, significantly surpassing the United States and South Korea. The article examines the evolution of patent strategies, the trend of technological shift from hardware to software, and the progress made by Chinese companies in patent quality, application scenarios, and global layout. It also points out that patent quantity does not equate to comprehensive technological leadership, emphasizing the challenges of manufacturing costs and quality, as well as the dual role of patents in industrial competition.
Robotic Systems Lab@leggedroboticsSignalEditorial score6363
MIT researchers have developed a new technique that helps generative AI models meet strict safety and task-specific constraints without sacrificing output quality. The method, called HardFlow, reformulates constraint satisfaction as a trajectory-optim, allowing models to explore more freely during generation while ensuring final compliance with hard constraints. It is tested on robotics, control, and computer vision tasks, consistently outperforming existing methods in constraint satisfaction and solution quality.
LeRobot has released a converted version of the Hy-Embodied-0.5-VLA-UMI model, ensuring full tensor compatibility and validation. The model supports a 16D dual-arm embodiment but cannot express 12D mobile-manipulator actions, indicating limitations in the current representation.
Article dates follow your selected timezone. Briefing editions use Hong Kong time (UTC+8).