Articles tagged with "deep-learning"
Robot learns adaptive walking on uneven terrain using deep learning
The article discusses a quadruped robot that has successfully learned to walk adaptively on slippery and uneven terrain solely through deep reinforcement learning in simulation, without relying on human-designed gaits or manual tuning. Traditional legged robot control methods depend on precise physical models and predefined motions, which often fail in unpredictable environments. To overcome these limitations, the researchers developed a structured training framework using a curriculum that gradually increases terrain difficulty—from flat ground to slopes, rough surfaces, low friction, and mixed conditions with sensor noise. This approach enables the robot to develop robust locomotion skills and adapt instinctively to new terrains. The robot’s control system features a hierarchical structure with a high-level neural network policy generating joint targets at 10 Hz, executed by a low-level controller at 100 Hz for stability. It uses proprioceptive inputs (joint angles, velocities, body orientation) and exteroceptive data from a simulated depth camera to perceive terrain features. Training employed proximal policy optimization with a multi-objective reward balancing
roboticsdeep-learningadaptive-walkingreinforcement-learningquadruped-robotterrain-navigationrobotic-control-systemsRobot Talk Episode 139 – Advanced robot hearing, with Christine Evers - Robohub
In Robot Talk Episode 139, Claire interviews Christine Evers, Associate Professor and Director of the Centre for Robotics at the University of Southampton, about advancing robot hearing capabilities. Evers’ research focuses on enabling robots to better understand their environment through sound by integrating insights from human auditory processes into deep-learning audio models. This bio-inspired approach aims to develop more compute-efficient and interpretable systems compared to traditional large-scale internet-trained models. The discussion highlights a shift toward embodied auditory intelligence, where robots gain a more nuanced and practical understanding of sound in real-world contexts. By embedding human auditory understanding into machine learning architectures, Evers’ work opens new possibilities for robots to interact with and interpret their surroundings more effectively. The episode is part of the Robot Talk podcast series, which covers developments in robotics, AI, and autonomous machines.
roboticsrobot-hearingmachine-listeningauditory-intelligencedeep-learningartificial-intelligenceautonomous-machinesChina's unmanned drones refuel autonomously under harsh conditions
Chinese researchers at Northwestern Polytechnical University have successfully demonstrated autonomous aerial refueling between two unmanned aerial vehicles (UAVs) under challenging conditions, marking a significant advancement in long-endurance drone operations. The test involved one UAV acting as a tanker and the other as a receiver, which autonomously located, tracked, and docked with the tanker despite visual obstacles such as glare and partial occlusion. This system relied on a machine vision approach using a dual-camera near-infrared setup and deep learning algorithms to detect and track the tanker’s refueling drogue with over 99% accuracy and centimeter-level positioning precision, enabling docking without human intervention during high-speed formation flight. This breakthrough has strategic implications, particularly for China’s heavy-duty Jiu Tian drone platform, which boasts a 7,000-kilometer range and the capacity to carry over 200 loitering munitions. While the drone’s unrefueled range does not reach the continental United States, autonomous aerial refueling
robotautonomous-dronesaerial-refuelingmachine-visiondeep-learningUAV-technologymilitary-aviationUS: World's smallest AI supercomputer that fits in a pocket unveiled
US startup Tiiny AI has unveiled the Tiiny AI Pocket Lab, officially recognized by Guinness World Records as the world’s smallest personal AI supercomputer. This pocket-sized device, resembling a power bank, can locally run large language models (LLMs) with up to 120 billion parameters without needing cloud connectivity, servers, or high-end GPUs. The Pocket Lab aims to reduce reliance on cloud infrastructure and GPUs, addressing sustainability concerns, rising energy costs, and privacy risks associated with cloud-based AI. By enabling advanced AI capabilities on a personal device, Tiiny AI seeks to make AI more accessible, private, and energy-efficient. Designed for a wide range of users including creators, developers, researchers, and students, the Pocket Lab supports complex AI tasks such as multi-step reasoning, deep context understanding, and secure processing of sensitive data—all while keeping data stored locally with bank-level encryption. It runs models between 10 billion and 100 billion parameters, covering over 80% of real-world AI tasks,
AIsupercomputerenergy-efficiencyedge-computingprivacydeep-learninglow-power-devicesEx-Tesla, Google, Nvidia leaders to build next-gen humanoid robot
UMA (Universal Mechanical Assistant) is a newly launched robotics intelligence company founded by former leaders from Tesla, Google DeepMind, Nvidia, and Hugging Face. The company’s mission is to develop next-generation humanoid and mobile robots capable of performing real-world industrial tasks alongside humans. The founding team includes experts who have contributed significantly to advances in deep learning, robot learning, and open-source AI, such as Rémi Cadene, Pierre Sermanet, Simon Alibert, and Robert Knight. UMA is focusing on creating compact, dual-armed mobile robots designed for environments like warehouses, assembly lines, hospitals, labs, and homes. UMA’s vision aligns with a broader market trend where robotics, rather than just generative AI or language models, will define the next era of artificial intelligence. Analysts project the global humanoid and mobile robotics market to reach $243 billion by 2035 and expand into a multi-trillion-dollar industry by 2050. This growth is driven by critical labor shortages—
roboticshumanoid-robotsAIindustrial-robotsautomationdeep-learningmobile-robotsAI-powered wearable turns everyday gestures into machine control
Researchers at the University of California San Diego have developed an AI-powered wearable system that accurately interprets natural arm gestures to control machines, even under intense motion disturbances such as running, riding in a car, or turbulent ocean conditions. This next-generation human–machine interface combines soft, stretchable sensors embedded in a thin electronic patch with a deep-learning model that filters out motion noise in real time, enabling reliable gesture recognition in real-world, high-motion environments. Unlike previous wearable gesture sensors that fail with excessive movement, this system maintains accuracy across a broad range of disturbances, making it suitable for diverse applications from medical rehabilitation to underwater robotics. The wearable device integrates motion and muscle sensors, a Bluetooth microcontroller, and a stretchable battery into a multilayered armband patch. It was rigorously tested in extreme scenarios, including simulated ocean conditions, where it demonstrated low-latency, precise control of a robotic arm despite disruptive motions. Originally inspired by military divers’ need for underwater robot control, the technology’s
robotwearable-technologyhuman-machine-interfacegesture-controldeep-learningmotion-sensorsunderwater-roboticsInterview with Zahra Ghorrati: developing frameworks for human activity recognition using wearable sensors - Robohub
In this interview, Zahra Ghorrati, a PhD candidate at Purdue University, discusses her research on developing scalable and adaptive deep learning frameworks for human activity recognition (HAR) using wearable sensors. Her work addresses the challenges posed by noisy, inconsistent, and uncertain data from wearable devices, aiming to create models that are computationally efficient, interpretable, and robust enough for real-world applications outside controlled lab environments. Unlike video-based recognition systems, wearable sensors offer privacy advantages and continuous monitoring capabilities, making them highly suitable for healthcare and long-term activity tracking. Ghorrati’s research has focused on a hierarchical fuzzy deep neural network that adapts to diverse HAR datasets by detecting simpler activities at lower levels and more complex ones at higher levels. By integrating fuzzy logic into deep learning, her model effectively handles uncertainty in sensor data, improving both robustness and interpretability. This approach also maintains low computational costs, enabling real-time recognition on wearable devices. Evaluations on multiple benchmark datasets show that her framework achieves competitive accuracy
robotwearable-sensorshuman-activity-recognitiondeep-learningIoThealthcare-technologysensor-data-analysisSelf-supervised learning for soccer ball detection and beyond: interview with winners of the RoboCup 2025 best paper award - Robohub
The article highlights the award-winning research on autonomous soccer ball detection by the SPQR team, who received the best paper award at RoboCup 2025 held in Salvador, Brazil. The team addressed a key challenge in robotic soccer: accurate ball detection under varying conditions. Traditional deep learning approaches require large labeled datasets, which are difficult and labor-intensive to produce for highly specific tasks like RoboCup. To overcome this, the researchers developed a self-supervised learning framework that reduces the need for manual labeling by leveraging pretext tasks that exploit the structure of unlabeled image data. Their method also incorporates external guidance from a pretrained object detection model (YOLO) to refine predictions from a general bounding box to a more precise circular detection around the ball. Deployed at RoboCup 2025, the new model demonstrated significant improvements over their 2024 benchmark, notably requiring less training data and exhibiting greater robustness to different lighting and environmental conditions. This adaptability is crucial given the variability of competition venues. The SPQR team
robotautonomous-robotsself-supervised-learningdeep-learningRoboCupsoccer-robotscomputer-visionNew Wi-Fi fingerprint system re-identifies people without devices
Italian researchers from La Sapienza University of Rome have developed a novel Wi-Fi fingerprinting system called WhoFi that can re-identify individuals based solely on how their bodies distort Wi-Fi signals, without requiring any carried devices like phones or wearables. By analyzing changes in Wi-Fi signal waveforms—specifically Channel State Information (CSI), which captures amplitude and phase alterations caused by a person’s physical presence—the system creates unique biometric identifiers. Using a transformer-based deep neural network, WhoFi achieved up to 95.5% accuracy in matching individuals across different Wi-Fi-covered spaces, significantly improving on previous methods such as the 75% accurate EyeFi system introduced in 2020. This approach offers a new dimension to surveillance and tracking, as Wi-Fi signals can penetrate walls and operate independently of lighting conditions, unlike cameras. While it may appear more privacy-conscious since it does not capture images, WhoFi raises concerns about passive tracking without consent. The technology builds on advances like the IEEE 802.
IoTWi-Fibiometric-identificationsignal-processingdeep-learningsurveillance-technologyChannel-State-InformationAnduril alums raise $24M Series A to bring military logistics out of the Excel spreadsheet era
Rune, a startup founded by former Anduril and military veterans, has raised $24 million in a Series A funding round to modernize military logistics through AI-enabled software. Co-founder David Tuttle highlighted that current U.S. military logistics rely heavily on outdated manual processes like Excel spreadsheets and whiteboards, which are insufficient for the scale and pace of modern warfare. Rune’s flagship product, TyrOS, aims to transform these processes into intelligent, predictive supply networks that optimize resources and support distributed operations, even in disconnected environments such as remote battlefields. TyrOS leverages deep learning models to forecast supply and demand for personnel, equipment, and other resources by analyzing hundreds of environmental and logistical variables. It also incorporates threat-informed routing and integrates generative AI for real-time "course of action" generation, helping commanders make informed decisions quickly. Despite advances in large language models, TyrOS maintains traditional mathematical optimization for precise logistical tasks like aircraft load planning. Its edge-first, cloud-capable but not cloud
IoTmilitary-logisticsAIdeep-learningsupply-chain-optimizationdefense-technologypredictive-analyticsWayve CEO Alex Kendall brings the future of autonomous AI to TechCrunch Disrupt 2025
At TechCrunch Disrupt 2025, taking place from October 27–29 at Moscone West in San Francisco, Alex Kendall, co-founder and CEO of Wayve, will be featured on an AI-focused panel discussing the future of autonomous AI. Kendall, who founded Wayve in 2017, has pioneered a new approach to autonomous driving that relies on embodied intelligence powered by deep learning and computer vision, rather than traditional handcrafted rules or maps. His work demonstrated that machines can interpret their environment and make real-time driving decisions without manual coding, marking a significant breakthrough in self-driving technology. Currently, Kendall is spearheading the development of AV2.0, a next-generation autonomous vehicle architecture designed for global scalability. His role as CEO involves integrating strategy, research, partnerships, and commercialization efforts to bring intelligent driving systems to market. With a strong academic background, including a PhD in Computer Vision and Robotics and recognition on Forbes 30 Under 30, Kendall brings a unique combination of scientific expertise
robotautonomous-vehiclesAIdeep-learningcomputer-visionembodied-intelligenceself-driving-systemsRobotic hand moves like magic, controlled by nothing but thought
Researchers at Carnegie Mellon University have achieved a breakthrough in noninvasive brain-computer interface (BCI) technology by enabling real-time control of a robotic hand’s individual fingers using only human thought. Utilizing electroencephalography (EEG) combined with a novel deep-learning decoding strategy, the system translates brain signals into precise finger movements without any muscle activity. Volunteers successfully performed multi-finger tasks, demonstrating the system’s ability to overcome traditional EEG spatial limitations and achieve fine motor control. Led by Professor Bin He, whose lab has pioneered several EEG-powered robotic controls, this innovation offers a risk-free, external alternative to invasive BCIs that require surgery. The technology holds significant promise for a broad range of users, including people with motor impairments or those recovering from injuries, by enhancing hand function and quality of life. Beyond medical rehabilitation, the system’s natural dexterity opens possibilities for everyday tasks like typing or manipulating small objects, potentially redefining how assistive devices integrate seamlessly as intuitive extensions of the human body
roboticsbrain-computer-interfacenoninvasive-BCIdeep-learningprostheticsassistive-technologyEEG-controlNew Insights for Scaling Laws in Autonomous Driving - CleanTechnica
The article from CleanTechnica discusses Waymo’s recent research into applying scaling laws—well-established in large language models (LLMs)—to autonomous driving, specifically in motion forecasting and planning. Waymo’s study leveraged an extensive internal dataset of 500,000 hours of driving, much larger than prior AV datasets, to investigate how increasing model size, training data, and compute resources impact AV performance. The findings reveal that, similar to LLMs, motion forecasting quality improves predictably following a power-law relationship with training compute. Additionally, scaling data and inference compute enhances the model’s ability to handle complex driving scenarios, and closed-loop planning performance also benefits from increased scale. These results mark a significant advancement by demonstrating for the first time that real-world autonomous vehicle capabilities can be systematically improved through scaling, providing a predictable path to better performance. This predictability applies not only to model training objectives and open-loop forecasting metrics but also to closed-loop planning in simulations, which more closely reflect real driving conditions.
robotautonomous-vehiclesAImotion-forecastingscaling-lawsdeep-learningWaymoHow to Make AI Faster and Smarter—With a Little Help From Physics
Rose Yu, an associate professor at UC San Diego, has pioneered the field of physics-guided deep learning by integrating physical principles, especially from fluid dynamics, into artificial neural networks. Her work began with addressing real-world problems like traffic congestion near the USC campus, where she modeled traffic flow as a diffusion process analogous to fluid flow, using graph theory to represent road networks and sensor data. This innovative approach allowed her to capture dynamic, time-evolving patterns in traffic, improving prediction accuracy beyond traditional static image-based deep learning methods. Yu’s research extends beyond traffic to other complex systems involving turbulence and spread phenomena. By embedding physics into AI models, she has accelerated simulations of turbulent flows to better understand hurricanes and developed tools to predict the spread of Covid-19. Her ultimate vision is to create AI Scientist, a collaborative framework where AI assistants, grounded in physical laws, partner with human researchers to enhance scientific discovery. This physics-informed AI approach promises to make AI both faster and smarter, enabling breakthroughs in diverse scientific and practical domains.
AIdeep-learningphysics-guided-learningtraffic-predictionturbulence-simulationdigital-lab-assistantsscientific-discovery