Articles tagged with "robot-training"
Physical Intelligence, Stripe veteran Lachy Groom’s latest bet, is building Silicon Valley’s buzziest robot brains
Physical Intelligence, co-founded by UC Berkeley associate professor Sergey Levine and led by Silicon Valley veteran Lachy Groom, is developing advanced robotic "brains" designed to enable robots to perform everyday tasks with human-like adaptability. Operating out of a modest San Francisco facility, the company uses off-the-shelf robotic arms priced around $3,500, emphasizing that sophisticated intelligence can compensate for relatively inexpensive hardware. Their approach involves training general-purpose robotic foundation models through continuous data collection from various environments, including warehouses, homes, and test kitchens. These models are then tested on tasks such as folding clothes or peeling vegetables, with the goal of enabling robots to generalize learned skills to new, unseen objects and challenges. The company’s work is likened to creating a "ChatGPT for robots," where the AI learns from diverse real-world interactions to improve its capabilities. Physical Intelligence’s experimental setup includes multiple robotic arms attempting mundane tasks, illustrating the iterative process of refining the models. Lachy Groom, a young entrepreneur with a background
roboticsartificial-intelligencerobotic-armsautomationmachine-learningrobotics-researchrobot-trainingNeo humanoid maker 1X releases world model to help bots learn what they see
Robotics company 1X, known for its Neo humanoid robots, has introduced a new physics-based AI model called the "world model" designed to help its bots better understand and learn from the real world. This model leverages video data combined with user prompts to enable Neo robots to acquire new skills beyond their initial training. While the company claims that Neo can transform any prompt into new actions, this capability is not immediate or all-encompassing; for example, Neo cannot instantly learn complex tasks like driving a car. Instead, the process involves capturing video linked to specific prompts, feeding this data into the world model, and then distributing the enhanced knowledge back to the network of Neo robots to improve their understanding and behavior over time. 1X is preparing to launch its Neo humanoids for home use, with pre-orders reportedly exceeding expectations, though the company has not disclosed exact shipping timelines or order quantities. According to founder and CEO Bernt Børnich, the world model marks a significant step toward
roboticsAIhumanoid-robotsmachine-learningrobotics-technologyautonomous-robotsrobot-trainingVideo: Humanoid robot obeys verbal commands, grabs Coke autonomously
Israel-based startup Mentee Robotics has demonstrated its Menteebot V3 humanoid robot autonomously responding to verbal commands, such as retrieving a can of Coke. The robot interprets spoken instructions, visually identifies the target object, navigates to it, grasps it, and returns to the user without human intervention. This capability is enabled by Mentee’s “Foundation Model,” which integrates language understanding, visual perception, navigation, and manipulation into a cohesive system. Training involves reinforcement learning in simulated environments, with skills transferred to real robots via Sim2Real techniques, allowing non-experts to teach robots naturally through speech and demonstration rather than coding. Founded in 2022 by Mobileye founder Prof. Amnon Shashua and AI experts, Mentee Robotics has raised over $40 million and employs about 70 people. On January 6 at CES, Mobileye announced plans to acquire Mentee in a deal valued up to $900 million, aiming to expand beyond autonomous vehicles into humanoid robotics
roboticshumanoid-robotautonomous-robotsmachine-learningAIrobot-traininghuman-robot-collaborationHumanoid robots to train in China’s first droid-friendly city zone
China is establishing its first robot-friendly urban demonstration zone in Shenzhen, Guangdong Province, to advance the integration of humanoid robots into everyday city life. Unveiled at the 2025 Guangdong-Hong Kong-Macao Greater Bay Area AI and Robotics Industry Conference, the Guangdong Embodied Intelligence Training Ground “1+1+N” framework aims to create coordinated training grounds for embodied intelligent robots. The framework includes a main provincial training ground, a management center, and the Shenzhen Embodied Intelligence Demonstration Zone, where robots will move from controlled environments to real-world urban settings. Multiple specialized sub-training grounds across cities and sectors will collaborate with local governments and industries to address practical challenges and foster innovation, accelerating the development and deployment of humanoid robots in manufacturing, public services, and other fields. Guangdong Province has become a national leader in robotics and AI, supported by strong manufacturing capabilities and targeted policies promoting innovation. Recent government measures include financial incentives and initiatives to boost core technology breakthroughs and expand application scenarios.
roboticshumanoid-robotsAI-integrationembodied-intelligencerobot-trainingsmart-cityGuangdong-robotics-zoneUsing generative AI to diversify virtual training grounds for robots - Robohub
Researchers at MIT’s CSAIL and the Toyota Research Institute have developed a novel “steerable scene generation” system to create diverse, realistic 3D digital environments—such as kitchens, living rooms, and restaurants—for training robots. Unlike traditional methods that rely on time-consuming real-world demonstrations or handcrafted simulations that often lack physical realism, this approach uses a diffusion model guided by Monte Carlo tree search (MCTS) to generate and refine scenes. Trained on over 44 million 3D rooms, the system places objects in new arrangements while ensuring physical accuracy, such as preventing object clipping, thereby producing lifelike environments that better mimic real-world physics. The key innovation lies in framing scene generation as a sequential decision-making process, where MCTS evaluates multiple scene variations to optimize for specific goals, such as physical realism or object diversity. This method allows the creation of complex scenes that surpass the complexity of the training data, exemplified by a restaurant scene containing up to 34 items—double the number
robotartificial-intelligencegenerative-AIrobot-trainingsimulation3D-modelingscene-generationFigure reaches $39B valuation in latest funding round
Figure, a San Jose-based humanoid robotics company, has secured its largest funding round to date, raising over $1 billion in a Series C round that values the company at $39 billion. The round was led by Parkway Venture Capital and included major investors such as Brookfield Asset Management, Nvidia, and Intel Capital. Since its founding in 2022, Figure has raised nearly $2 billion, reflecting strong investor interest in robots designed to work alongside humans in industrial environments like warehouses and factories. The new funding will be used to expand Figure’s fleet of humanoid robots, develop infrastructure to speed up robot training, and enhance advanced data collection efforts. Despite its rapid growth and high valuation, Figure has taken legal steps to control its stock distribution, issuing cease-and-desist letters to unauthorized secondary market brokers. CEO Brett Adcock has previously described Figure as the most “sought-after” private stock earlier in 2025.
roboticshumanoid-robotswarehouse-automationrobot-trainingAI-roboticsindustrial-robotsrobotics-fundingWhy Runway is eyeing the robotics industry for future revenue growth
Runway, a New York-based company known for its AI-powered video and photo generation tools built over the past seven years, is now targeting the robotics industry as a new avenue for revenue growth. The company’s advanced world models—large language models that simulate realistic versions of the real world—have attracted interest from robotics and self-driving car companies seeking scalable and cost-effective training simulations. Runway’s co-founder and CTO, Anastasis Germanidis, explained that while the company initially focused on creative and entertainment applications, inbound requests from robotics firms revealed broader use cases for their technology beyond entertainment. Robotics companies are leveraging Runway’s models to create highly specific training simulations that are difficult, costly, and time-consuming to replicate in real-world environments. These simulations allow for controlled testing of different actions and scenarios without altering other environmental variables, providing valuable insights into outcomes that physical testing cannot easily achieve. Rather than developing separate models for robotics and autonomous vehicles, Runway plans to fine-tune its existing models and is
roboticsAI-simulationself-driving-carsrobot-trainingvisual-generating-toolsrobotics-industrymachine-learningRobot Dog Gets Trained To Backflip
Boston Dynamics has demonstrated their quadrupedal robot, Spot, performing backflips as part of a training regimen aimed at enhancing the robot's balance and recovery capabilities. This exercise is designed to push the limits of Spot's agility and stability, showcasing advancements in robotic movement and control. The backflip training highlights Boston Dynamics' ongoing efforts to improve the robot's ability to handle dynamic and challenging motions, which could translate to better performance in real-world applications requiring agility and resilience. This development underscores the company's commitment to refining robotic mobility through complex physical maneuvers.
robotroboticsBoston-Dynamicsquadrupedal-robotrobot-dogrobot-trainingrobot-balanceHow to train generalist robots with NVIDIA's research workflows and foundation models - The Robot Report
NVIDIA researchers are advancing scalable robot training by leveraging generative AI, world foundation models (WFMs), and synthetic data generation workflows to overcome the traditional challenges of collecting and labeling large datasets for each new robotic task or environment. Central to this effort is the use of WFMs like NVIDIA Cosmos, which are trained on millions of hours of real-world data to predict future states and generate video sequences from single images. This capability enables rapid, high-fidelity synthetic data generation, significantly accelerating robot learning and reducing development time from months to hours. Key components of NVIDIA’s approach include DreamGen, a synthetic data pipeline that creates diverse and realistic robot trajectory data with minimal human input, and GR00T models that facilitate generalist skill learning across varied tasks and embodiments. The DreamGen pipeline involves four main steps: post-training a world foundation model (e.g., Cosmos-Predict2) on a small set of real demonstrations, generating synthetic photorealistic robot videos from image and language prompts, extracting pseudo-actions
roboticsartificial-intelligencesynthetic-data-generationNVIDIA-Isaacfoundation-modelsrobot-trainingmachine-learning