Articles tagged with "robot-learning"
Skild AI brain lets robots watch videos to master everyday tasks
Skild AI has developed Skild Brain, an innovative AI model that enables robots to learn everyday tasks by watching human videos and simulations, significantly reducing the need for extensive robot-specific training data. This approach allows robots to perform a variety of activities such as opening doors, watering plants, assembling boxes, cooking, and navigating challenging terrains with high precision and adaptability. Unlike traditional robotics methods that rely heavily on teleoperation and limited datasets, Skild Brain leverages large-scale human video data from the internet, bridging the “embodiment gap” by mapping human actions onto diverse robot morphologies—including humanoids, quadrupeds, and mobile manipulators—without retraining. Founded in 2023 by experts in self-supervised and adaptive robotics, Skild AI addresses the critical data bottleneck in robotics by pre-training its foundation model on vast human video datasets and physics simulations. Skild Brain is omni-bodied and resilient, capable of adapting in real time to unpredictable conditions such as limb loss or payload changes without
roboticsartificial-intelligencehumanoid-robotsrobot-learningadaptive-roboticsfoundation-modelsrobot-automationVideo: Humanoid's Alpha reenacts Rowan Atkinson's Love Actually scene
Humanoid, a UK-based robotics firm, released a holiday video featuring its AI-powered humanoid robot HMND 01 Alpha Bipedal reenacting the iconic gift-wrapping scene from the 2003 Christmas film "Love Actually," originally performed by Rowan Atkinson. The one-and-a-half-minute video humorously mirrors the original scene’s tone, showing Alpha as a shop clerk who offers to gift-wrap a small robot. The robot’s exaggerated and increasingly elaborate wrapping, complete with festive decorations, highlights both comedic elements and the robot’s dexterity and learning process. The scene culminates with Alpha recognizing the excess and discreetly replacing the over-wrapped package with a simpler one, underscoring themes of adaptation, judgment, and the trial-and-error nature of teaching humanoid robots. HMND 01 Alpha, standing nearly six feet tall and weighing about 200 pounds, was built in just five months and achieved stable bipedal walking within 48 hours of assembly. The robot has logged
roboticshumanoid-robotAI-robotbipedal-robotrobot-dexterityrobotics-in-retailrobot-learningTeaching 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-adaptationChinese model helps humanoid robots adapt to tasks without training
Researchers from Wuhan University have developed a novel framework called the recurrent geometric-prior multimodal policy (RGMP) to enhance humanoid robots' ability to manipulate objects with human-like adaptability and minimal training. Current humanoid robots excel at specific tasks but struggle to generalize when objects change shape, lighting varies, or when encountering tasks they were not explicitly trained for. RGMP addresses these limitations by incorporating two key components: the Geometric-Prior Skill Selector (GSS), which helps the robot analyze an object's shape, size, and orientation to select the appropriate skill, and the Adaptive Recursive Gaussian Network (ARGN), which models spatial relationships and predicts movements efficiently with far fewer training examples than traditional deep learning methods. Testing showed that robots using RGMP achieved an 87% success rate on novel tasks without prior experience, demonstrating a significant improvement over existing diffusion-policy-based models, with about five times greater data efficiency. This advancement could enable humanoid robots to perform a wider range of tasks in dynamic environments such
roboticshumanoid-robotsrobot-learningdata-efficient-roboticsrobotic-manipulationAI-in-roboticsrobotic-skill-adaptationRobot Talk Episode 130 – Robots learning from humans, with Chad Jenkins - Robohub
In the Robot Talk Episode 130 podcast, Claire interviews Chad Jenkins, a Professor of Robotics and Electrical Engineering and Computer Science at the University of Michigan, about how robots can learn from humans to better assist in daily tasks. Jenkins’ research focuses on robot learning from demonstration, human-robot interaction, dexterous mobile manipulation, and robot perception. Notably, he founded the Robotics Major Degree Program at the University of Michigan in 2022 and received the 2024 ACM/CMD-IT Richard A. Tapia Achievement Award for his contributions to scientific scholarship, civic science, and diversity in computing. The episode highlights the intersection of robotics and human collaboration, emphasizing how robots can be taught by observing human actions to improve their functionality and integration into everyday life. This discussion fits within the broader context of the Robot Talk podcast series, which explores advancements in robotics, AI, and autonomous machines, featuring experts from various fields. The episode also connects to related topics such as robotic applications in smart cities, museum
roboticsrobot-learninghuman-robot-interactionautonomous-machinesrobot-perceptionmicrorobotsrobotic-systemsWhat’s coming up at #IROS2025? - Robohub
The 2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS 2025) will take place from October 19 to 25 in Hangzhou, China. The event features a comprehensive program including plenary and keynote talks, workshops, tutorials, forums, competitions, and a debate. The three plenary talks scheduled for October 21-23 will cover topics such as humanoid and quadruped mobility in real-world applications (Marco Hutter), autonomous aerial manipulation for physically intelligent flying robots (Hyoun Jin Kim), and the integration of physical robots with artificial general intelligence agents (Song-Chun Zhu). Keynote presentations are organized under eleven thematic areas, highlighting cutting-edge research and developments in robotics. These areas include Rehabilitation & Physically Assistive Systems, Bio-inspired Robotics, Soft Robotics, AI and Robot Learning, Perception and Sensors, Human-Robot Interaction, Embodied Intelligence, Medical Robots, and Field Robotics. Notable topics include advancements in legged robots and
roboticssoft-roboticsAIhumanoid-robotswearable-robotsrobot-learningautonomous-systemsDyna Robotics closes $120M funding round to scale robotics foundation model - The Robot Report
Dyna Robotics Inc. has secured $120 million in a Series A funding round to accelerate the development of its next-generation robotics foundation model aimed at creating general-purpose robots for commercial environments. Building on its earlier $23.5 million seed round and the launch of its DYNA-1 model, the company emphasizes that its proprietary foundation model enables robots to achieve over a 99% success rate during continuous 24-hour operation. The model’s strength lies in its ability to generalize across diverse environments—such as hotels, restaurants, laundromats, and gyms—allowing robots to function effectively out of the box without additional data, and to improve rapidly through on-the-job learning. Dyna’s approach centers on designing foundation models that combine generalization with high performance, enabling robots to master a wide range of manipulation skills and adapt to complex tasks through continuous learning. The company’s leadership team, including co-founders Lindon Gao and York Yang and former Google DeepMind researcher Jason Ma, brings extensive
roboticsfoundation-modelsAI-roboticsgeneral-purpose-robotsrobot-learningcommercial-robotsrobotics-fundingBoston Dynamics’ Atlas Gets a Brain Upgrade
Boston Dynamics has significantly upgraded its humanoid robot Atlas by integrating a Large Behavior Model (LBM), enabling the robot to learn complex human actions from extensive datasets rather than relying on traditional hand-coded instructions. This advancement allows Atlas to perform a variety of tasks with notable dexterity, such as moving baskets, transferring objects, placing items on shelves, and manipulating different shapes. The robot also demonstrates resilience by continuing its work despite attempts to disrupt it. This shift to LBMs marks a move toward creating truly general-purpose humanoid robots capable of adapting quickly to real-world environments and tasks. By leveraging large-scale learning models, Atlas can exhibit more flexible and autonomous behavior, potentially broadening its applications beyond pre-programmed routines. Boston Dynamics’ decision to withhold Atlas from the inaugural Robot Olympics in China reflects their focus on refining this sophisticated capability before public competition.
robotBoston-Dynamicshumanoid-robotLarge-Behavior-Modelrobotics-AIAtlas-robotrobot-learningFieldAI raises $405M to build universal robot brains
FieldAI, a robotics AI company, announced a $405 million funding raise to develop universal "robot brains" capable of controlling diverse physical robots across varied real-world environments. The latest funding round, including a $314 million tranche co-led by Bezos Expedition, Prysm, and Temasek, adds to backing from investors such as Khosla Ventures and Intel Capital. FieldAI’s core innovation lies in its "Field Foundation Models," which integrate physics-based understanding into embodied AI—AI that governs robots physically navigating environments—enabling robots to quickly learn, adapt, and manage risk and safety in new settings. This physics-informed approach contrasts with traditional AI models that often lack risk awareness, making FieldAI’s robots better suited for complex and potentially hazardous environments. Founder and CEO Ali Agha emphasized that their goal is to create a single, general-purpose robot brain that can operate across different robot types and tasks, with a built-in confidence measure to assess decision reliability and manage safety thresholds. Agha’s decades
robotartificial-intelligenceembodied-AIrobotics-safetyrobot-learningAI-modelsrobotics-technologyTop 10 robotics developments of July 2025 - The Robot Report
In July 2025, the robotics industry saw significant advancements, funding milestones, and strategic partnerships, as highlighted by The Robot Report's top 10 articles. MIT introduced a novel 3-in-1 training tool that simplifies robot learning by allowing robots to learn tasks through demonstration, either via remote control, physical guidance, or observation. NEURA Robotics partnered with HD Hyundai to develop specialized quadruped and humanoid robots tailored for the demanding shipbuilding sector, showcasing the expanding versatility of cognitive robots. Meanwhile, RealSense spun off from Intel to operate independently with $50 million in funding, focusing on 3D vision technologies for robotics. Several companies secured substantial funding to accelerate innovation: Augmentus raised $11 million to develop no-code robotics programming solutions aimed at reducing complexity for manufacturers; Unitree Robotics achieved unicorn status with a Series C round valuing the company at approximately $1.7 billion, reflecting growing interest in legged robots; and Galbot secured $153 million to commercialize its
roboticsrobot-learningindustrial-robotsquadruped-robotshumanoid-robotsrobot-fundingrobotic-technologyInterview with Kate Candon: Leveraging explicit and implicit feedback in human-robot interactions - Robohub
In this interview, Kate Candon, a PhD student at Yale University, discusses her research on improving human-robot interaction by leveraging both explicit and implicit feedback. Traditional robot learning often relies on explicit feedback, such as simple "good job" or "bad job" signals from a human teacher who is not actively engaged in the task. However, Candon emphasizes that humans naturally provide a range of implicit cues—like facial expressions, gestures, or subtle actions such as moving an object away—that convey valuable information without additional effort. Her current research aims to develop a framework that combines these implicit signals with explicit feedback to enable robots to learn more effectively from humans in natural, interactive settings. Candon explains that interpreting implicit feedback is challenging due to variability across individuals and cultures. Her initial approach focuses on analyzing human actions within a shared task to infer appropriate robot responses, with plans to incorporate visual cues such as facial expressions and gestures in future work. The research is tested in a pizza-making scenario, chosen for
robothuman-robot-interactionimplicit-feedbackexplicit-feedbackinteractive-agentsrobot-learningAIMIT’s 3-in-1 training tool eases robot learning
MIT engineers have developed a novel three-in-one training interface that allows robots to learn new tasks through any of three common demonstration methods: remote control (teleoperation), physical manipulation (kinesthetic training), or by observing a human perform the task (natural teaching). This handheld, sensor-equipped tool can attach to many standard robotic arms, enabling users to teach robots in whichever way best suits the task or user preference. The interface was tested on a collaborative robotic arm by manufacturing experts performing typical factory tasks, demonstrating increased flexibility in robot training. This versatile demonstration interface aims to broaden the range of users who can effectively teach robots, potentially expanding robot adoption beyond manufacturing into areas like home care and healthcare. For example, one person could remotely train a robot to handle hazardous materials, another could physically guide the robot in packaging, and a third could demonstrate drawing a logo for the robot to mimic. The research, led by MIT’s Department of Aeronautics and Astronautics and CSAIL, was presented at the IEEE I
roboticsrobot-learninghuman-robot-interactioncollaborative-robotsrobot-training-toolsMIT-roboticsintelligent-robotsTRI: pretrained large behavior models accelerate robot learning
The Toyota Research Institute (TRI) has advanced the development of Large Behavior Models (LBMs) to accelerate robot learning, demonstrating that a single pretrained LBM can learn hundreds of tasks and acquire new skills using 80% less training data. LBMs are trained on large, diverse datasets of robot manipulation, enabling general-purpose robots to perform complex, long-horizon behaviors such as installing a bike rotor. TRI’s study involved training diffusion-based LBMs on nearly 1,700 hours of robot data and conducting thousands of real-world and simulation rollouts, revealing that LBMs consistently outperform policies trained from scratch, require 3-5 times less data for new tasks, and improve steadily as more pretraining data is added. TRI’s LBMs use a diffusion transformer architecture with multimodal vision-language encoders and a transformer denoising head, processing inputs from wrist and scene cameras, proprioception, and language prompts to predict short action sequences. The training data combines real-world teleoperation data,
roboticslarge-behavior-modelsrobot-learningpretrained-modelsToyota-Research-Instituteautonomous-robotsembodied-AICongratulations to the #ICRA2025 best paper award winners - Robohub
The 2025 IEEE International Conference on Robotics and Automation (ICRA), held from May 19-23 in Atlanta, USA, announced its best paper award winners and finalists across multiple categories. The awards recognized outstanding research contributions in areas such as robot learning, field and service robotics, human-robot interaction, mechanisms and design, planning and control, and robot perception. Each category featured a winning paper along with several finalists, highlighting cutting-edge advancements in robotics. Notable winners include "Robo-DM: Data Management for Large Robot Datasets" by Kaiyuan Chen et al. for robot learning, "PolyTouch: A Robust Multi-Modal Tactile Sensor for Contact-Rich Manipulation Using Tactile-Diffusion Policies" by Jialiang Zhao et al. for field and service robotics, and "Human-Agent Joint Learning for Efficient Robot Manipulation Skill Acquisition" by Shengchent Luo et al. for human-robot interaction. Other winning papers addressed topics such as soft robot worm behaviors, robust sequential task solving via dynamically composed gradient descent, and metrics-aware covariance for stereo visual odometry. The finalists presented innovative work ranging from drone detection to adaptive navigation and assistive robotics, reflecting the broad scope and rapid progress in the robotics field showcased at ICRA 2025.
roboticsrobot-learninghuman-robot-interactiontactile-sensorsrobot-automationsoft-roboticsrobot-navigation