Articles tagged with "robot-manipulation"
Robohub highlights 2025 - Robohub
The article "Robohub highlights 2025" provides a comprehensive review of notable contributions and activities featured on Robohub throughout the year. It highlights a variety of research and discussions from global experts in robotics and AI, including innovative frameworks for robot manipulation learned solely from language instructions, as presented by Jiahui Zhang and Jesse Zhang, and RobustDexGrasp, a novel grasping framework introduced by Hui Zhang at CoRL 2025. The article also covers insightful interviews and podcasts, such as conversations with Heather Knight on integrating performing arts methods into robotics, and Professor Marynel Vázquez on human-robot interactions and social navigation by robots. Further, the summary touches on advancements in reliable controller design under uncertain environments (IJCAI 2025), reinforcement learning guided by social and ethical norms, and scalable deep learning for human activity recognition using wearable sensors. It also features updates from RoboCup 2025, including award-winning research, AI applications in the Small Size League, and the Robo
roboticsrobot-manipulationhuman-robot-interactionreinforcement-learningAI-in-roboticsRoboCuprobot-graspingTeaching robot policies without new demonstrations: interview with Jiahui Zhang and Jesse Zhang - Robohub
The article discusses the ReWiND framework introduced by Jiahui Zhang, Jesse Zhang, and colleagues in their CoRL 2025 paper, which enables robots to learn manipulation policies for new language-specified tasks without requiring new demonstrations. ReWiND operates in three phases: first, it learns a dense reward function from a small set of demonstrations in the deployment environment by predicting per-frame progress toward task completion. A novel video rewind augmentation technique synthetically generates sequences simulating progress and failure to improve reward model accuracy and generalization. Second, the reward function is used to relabel demonstration data with dense rewards, enabling offline reinforcement learning to pre-train a policy. Finally, the pre-trained policy is fine-tuned online on unseen tasks using the frozen reward function as feedback, allowing continual adaptation without additional demonstrations. The researchers validated ReWiND through experiments in both the MetaWorld simulation and a real-world robotic setup (Koch). They focused on the reward model’s ability to generalize to unseen tasks
roboticsrobot-learningreinforcement-learninglanguage-guided-roboticsrobot-manipulationreward-function-learningpolicy-adaptationChina AI helps humanoid robots handle more objects with less training
Researchers at Wuhan University in China have developed a novel AI framework called RGMP (recurrent geometric-prior multimodal policy) to enhance humanoid robot manipulation capabilities. RGMP integrates geometric reasoning with efficient learning to improve grasping accuracy and enable robots to handle a wider variety of objects and more complex tasks with significantly less training data. Unlike many existing data-driven methods that require large datasets and struggle to generalize beyond familiar environments, RGMP achieves 87 percent generalization accuracy and is five times more data-efficient than leading diffusion-based models. The framework consists of two main components: the Geometric-prior Skill Selector (GSS), which chooses appropriate actions based on object shape and task needs using geometric rules, and the Adaptive Recursive Gaussian Network (ARGN), which models spatial memory over time to improve learning from limited examples. The team tested RGMP on both humanoid and desktop dual-arm robots using a dataset of 120 demonstration trajectories, comparing its performance against state-of-the-art models like ResNet50
roboticshumanoid-robotsAI-frameworkrobot-manipulationgeometric-reasoningmachine-learningdata-efficient-learningCoRL2025 – RobustDexGrasp: dexterous robot hand grasping of nearly any object - Robohub
The article "CoRL2025 – RobustDexGrasp: dexterous robot hand grasping of nearly any object" discusses the significant challenges and advancements in enabling robotic hands to achieve human-like dexterity in grasping diverse objects. Human hands possess over 20 degrees of freedom, allowing them to perform complex and adaptive manipulations effortlessly. In contrast, most robots currently use simple grippers that lack this versatility, limiting their functionality in unstructured environments. The fundamental challenge lies in the high-dimensional control complexity of dexterous hands, the need to generalize grasping strategies across a wide variety of object shapes, and the difficulty of perceiving object geometry accurately using only monocular vision without detailed 3D models. To overcome these obstacles, the authors introduce RobustDexGrasp, a novel framework designed to enable robust dexterous grasping. Their approach employs a teacher-student curriculum in reinforcement learning: a teacher policy trained in simulation with full object and tactile information guides a student policy that learns
roboticsdexterous-robot-handrobotic-graspingrobot-manipulationrobotic-control-systemsrobotic-perceptionrobotic-dexterityFrom teleoperation to autonomy: Inside Boston Dynamics' Atlas training
In Episode 212 of The Robot Report Podcast, Boston Dynamics’ VP of robotics research, Scott Kuindersma, discussed the development of large behavior models (LBMs) for the Atlas humanoid robot. The team collected 20 hours of teleoperation data to train these LBMs, which enable Atlas to generalize manipulation tasks such as bi-manual operations, including picking and placing parts for the Spot quadruped robot. The development process involved data collection, annotation, model training, and evaluation, with a strong emphasis on combining simulation data and human demonstration data. Boston Dynamics plans to further test Atlas in Hyundai facilities and leverage AI-driven advancements to improve humanoid manipulation and dynamic behaviors. The episode also covered recent robotics industry news, including Serve Robotics’ acquisition of Voysys’ assets to enhance its autonomous delivery fleet with low-latency video streaming for remote monitoring and teleoperation. Zoox, an Amazon subsidiary, launched a free robotaxi service on the Las Vegas Strip, with plans to expand testing
roboticsBoston-DynamicsAtlas-robotteleoperationautonomous-robotsAI-in-roboticsrobot-manipulationAgility Robotics explains how to train a whole-body control foundation model - The Robot Report
Agility Robotics has developed a whole-body control foundation model for its Digit humanoid robot, designed to enable safe, stable, and versatile task execution in complex, human-centric environments. This model acts like a "motor cortex," integrating signals from different control layers to manage voluntary movements and fine motor skills. It is implemented as a relatively small LSTM neural network with fewer than one million parameters, trained extensively in NVIDIA’s Isaac Sim physics simulator. Remarkably, the model transfers directly from simulation to the real world without additional training, allowing Digit to perform tasks such as walking, grasping, and manipulating heavy objects with high precision and robustness to disturbances. The model can be prompted using various inputs, including dense spatial objectives and large language models, enabling Digit to execute complex behaviors like grocery shopping demonstrated at NVIDIA’s GTC event. Agility Robotics aims to provide an intuitive interface for humanoid robots similar to fixed-base robots, where users specify desired end-effector poses and the robot autonomously positions itself accordingly.
roboticshumanoid-robotswhole-body-controlneural-networksAI-in-roboticsrobot-manipulationAgility-RoboticsBoston Dynamics and TRI use large behavior models to train Atlas humanoid - The Robot Report
Boston Dynamics, in collaboration with Toyota Research Institute (TRI), is advancing the development of large behavior models (LBMs) to enhance the capabilities of its Atlas humanoid robot. Recognizing that humanoid robots must competently perform a wide range of tasks—from manipulating delicate objects to handling heavy items while maintaining balance and avoiding obstacles—Boston Dynamics is focusing on creating AI generalist robots. Their approach involves training end-to-end, language-conditioned policies that enable Atlas to execute complex, long-horizon manipulation tasks by leveraging its full-body mobility, including precise foot placement, crouching, and center-of-mass shifts. The development process involves four key steps: collecting embodied behavior data via teleoperation on both real hardware and simulations; processing and annotating this data for machine learning; training neural network policies across diverse tasks; and evaluating performance to guide further improvements. To maximize task coverage, Boston Dynamics employs a teleoperation system combining Atlas’ model predictive controller with a custom VR interface, enabling the robot to perform tasks
roboticshumanoid-robotsBoston-DynamicsAI-in-roboticsmachine-learningrobot-manipulationautomationBoston Dynamics’ humanoid robot handles annoying co-worker gracefully
Boston Dynamics, in collaboration with the Toyota Research Institute (TRI), has made significant advancements in its humanoid robot Atlas by developing a Large Behavior Model (LBM). This new system, trained on extensive datasets of human actions, enables Atlas to understand, generate, and adapt complex human behaviors in real-world settings without the need for laborious hand-coding. A recently released video demonstrates Atlas performing precise human-like tasks such as picking up and transferring objects, walking, crouching, and organizing items, albeit at a somewhat slow pace. Notably, the robot maintained focus and completed its tasks despite repeated disturbances from a human, showcasing improved robustness and adaptability. The integration of LBMs represents a paradigm shift in robotics, allowing new skills to be added quickly through human demonstrations rather than traditional programming. According to Boston Dynamics and TRI executives, this approach enhances generalization across long-horizon manipulation tasks and whole-body control, potentially transforming how robots operate in existing environments. The project, co-led by Scott Kuinders
roboticshumanoid-robotBoston-DynamicsAtlas-robotrobot-behavior-modelAI-in-roboticsrobot-manipulationFieldAI raises $405M to scale 'physics first' foundation models for robots - The Robot Report
FieldAI, a Mission Viejo, California-based robotics company, has raised $405 million through two consecutive funding rounds to accelerate its global expansion and product development. The company plans to double its workforce by the end of the year as it advances its work in locomotion and manipulation for autonomous robots. FieldAI’s technology centers on its proprietary Field Foundation Models (FFMs), a novel class of AI models specifically designed for embodied intelligence in robotics. Unlike standard vision or language models adapted for robotics, FFMs are built from the ground up to handle uncertainty, risk, and physical constraints in dynamic, unstructured environments without relying on prior maps, GPS, or fixed paths. FieldAI’s FFMs enable robots to safely and reliably perform complex tasks in diverse real-world industrial settings such as construction, energy, manufacturing, urban delivery, and inspection. This approach allows robots to dynamically adapt to new and unexpected conditions without manual programming, marking a significant breakthrough in robotics AI. The company’s investors include prominent names such as
roboticsartificial-intelligenceautonomous-robotsField-Foundation-Modelsindustrial-robotsrobot-locomotionrobot-manipulationSwiss robot dog can now pick up and throw a ball accurately like humans
ETH Zurich’s robotic dog ANYmal, originally designed for autonomous operation in challenging environments, has been enhanced with a custom arm and gripper, enabling it to pick up and throw objects with human-like accuracy. The robot’s advanced actuators and integrated sensors allow it to navigate complex terrain while maintaining stability and situational awareness. Unlike traditional factory robots, ANYmal is built to handle unpredictable outdoor conditions, making it suitable for tasks such as industrial inspection, disaster response, and exploration. The research team, led by Fabian Jenelten, trained ANYmal using reinforcement learning within a highly realistic virtual environment that simulated real-world physics. This approach, known as sim-to-real transfer, allowed the robot to practice millions of throws safely and ensured its skills transferred effectively to real-world scenarios. In testing, ANYmal successfully picked up and threw various objects—including balls, bottles, and fruit—across different surfaces and environmental challenges, such as wind and uneven ground, demonstrating adaptability and precise control without pre-programmed steps. This
roboticsautonomous-robotsreinforcement-learninglegged-robotsrobot-manipulationsim-to-real-transferrobot-perception