Sunday Robotics Unveils Memo: A Household Robot Trained on 10 Million Real-World Tasks

The age of household robots might finally be here: Meet Memo

Serving tech enthusiasts for over 25 years. TechSpot means tech analysis and advice you can trust. Looking ahead: Home robots are no longer limited to tightly scripted routines or one-off public demonstrations as they progress from prototype to pilot programs. With one company's robot approaching its beta test, a key question looms: How close are we to a future where robots can seamlessly handle everyday household chores? In a Mountain View kitchen, a wheeled robot named Memo methodically prepares an espresso, illustrating the technical progress underway at Sunday Robotics. The machine, roughly the size of a compact dishwasher, features two articulated arms with pincer-like grippers, a height-adjustable central column, and a display that mimics a cartoon face topped with a stylized red cap. Unlike humanoid robots designed to imitate human movement, Memo relies exclusively on wheels for mobility. Memo's coffee-making demonstration unfolds with measured precision. After positioning itself at the countertop, the robot uses one pincer to fill the portafilter with ground coffee, tamps it down using force-control algorithms, locks the filter into an espresso machine, places a cup, initiates the brewing cycle, and ultimately delivers the finished espresso to its recipient. Each step - mundane for a human - reveals substantial technical hurdles for robotics. Object recognition in cluttered spaces, variable grip reliability, and subtle force modulation are all essential; even small missteps could cause spills or equipment damage. Tony Zhao and Cheng Chi, Sunday Robotics' founding engineers, see vertical integration as key to overcoming the most formidable challenge in domestic robotics: adaptability. Their team not only designs Memo's hardware but also develops the AI models required for nuanced control and learning in ever-changing environments. To facilitate dexterous manipulation, Sunday uses a novel data-acquisition system: remote operators wear $400 sensor-equipped gloves that mirror the movements of Memo's mechanical hands while performing real household chores. These gloves allow the company to capture direct measurements of human handling strategies including grip strength, finger placement, and motion trajectories. As operators complete tasks, the data feeds into Memo's training pipeline, which fuses glove telemetry, vision input, and proprioceptive sensor readings to refine its manipulation models. This approach bypasses more conventional teleoperation systems in which operators control robots via cameras and joysticks. With glove-based training, each subtle aspect of manipulation - how a glass is grasped, tilted, or rotated - is mapped directly onto Memo's hardware. The result is more natural motion and greater effectiveness when dealing with unfamiliar objects and messy real-world setups. In one test, Memo gripped two differently sized glasses using distinct parts of a single hand, demonstrating a nuanced grasping ability enabled by data that more closely reflects human dexterity than traditional teleoperation or reinforcement-learning methods based on randomized trials. Roboticists, including UC Berkeley's Ken Goldberg, note that capturing rich, glove-derived manipulation data is an innovative step that could accelerate progress in real-world robotic adaptability. Such data is particularly valuable as robots transition from predictable factory settings to the dynamic, cluttered environments of everyday homes. Recent advancements have also allowed robots to leverage large language models, offering new ways for them to interpret tasks and respond intelligently to spoken or written instructions. However, the absence of a vast, shared data repository - a "robotics internet," as Zhao describes it - remains a significant roadblock. Instead, progress depends on collecting vast, diverse datasets that reflect the realities of home life. Memo's development is unfolding in a landscape where several companies are racing to bring robotics into the domestic sphere. Startups like Physical Intelligence, Skild, and Generalist are pursuing flexible, adaptive training approaches, while 1x has introduced a teleoperation-assisted humanoid home robot. Sunday Robotics differentiates itself through a tightly integrated hardware - software stack - a point frequently cited by investors and industry observers. Sarah Guo, founder of Conviction, highlights the team's blend of Tesla and Google DeepMind veterans, while Benchmark's Eric Vishria points to Sunday's emphasis on practical, real-world deployment. The coming year will see Memo tested in real homes, a critical evaluation of technical reliability and user satisfaction as the robot navigates children, pets, clutter, and the incomplete instructions typical of most households. Data from this pilot phase will guide Memo's path toward broader adoption. Early enthusiasts - much like the pioneers of personal computing - are expected to shape the robot's evolution and influence its learning curve. Sunday Robotics also plans to let users teach Memo new tasks directly, adding a layer of user-driven customization that reflects broader trends in human - AI interaction.

Household humanoid robot trained on 10 million chores unveiled in US

Eric Vishria, General Partner at Benchmark, echoed the significance of real-world data, calling it essential for robots that can operate outside controlled demonstration settings. "The promise of AI robotics isn't back-flipping or dancing demos, but robots that work in messy, real-world situations. To have those, we need real-world training data," he noted, describing the current industry data volume as only a fraction of what is required. Memo's physical design diverges from the push toward fully humanoid robots. Instead of legs, Memo uses a wheeled base for mobility and balance, with a central column that can raise or lower its torso to reach different heights. This structure keeps the robot stable even during power loss, avoiding the risk of falls. The robot features a glossy white body with two arms, a friendly cartoon-like face with long, button-like eyes, and interchangeable colorful baseball caps. The look leans toward a retro-futuristic aesthetic, a little bit like Baymax from the Movie Big Hero 6 or early Nintendo-era hardware rather than hyper-realistic humanoids. The company describes Memo's silicone-clad exterior as soft and approachable, designed to blend into homes rather than stand out as industrial machinery. The company will begin accepting applications for Memo's Founding Family Beta on November 19, 2025. Fifty households will be selected as early adopters, each receiving a numbered unit and direct support as Sunday continues refining capabilities ahead of a wider release.

Sunday wants to put a robot in every home, beginning with the launch of Memo - SiliconANGLE

Sunday wants to put a robot in every home, beginning with the launch of Memo Sunday, an artificial intelligence robotics startup founded by Stanford Ph.D. roboticists, launched today to introduce Memo: a household robot capable of doing everyday chores. The new robot is built for homes, with safety and stability as a priority. Founders Tony Zhao, chief executive, and Cheng Chi, chief technology officer, started Sunday in a garage, working 24/7 with 3D printers to develop and build the robot's hardware and appearance. "We built Memo to give people back time for what matters, with the safety needs for any family in mind," said Zhao. "This is a turning point for home robotics." Memo diverges from the current rush to build humanoid-style robots in that it has no legs, which removes one of the biggest problems with humanoid robots: modelling balance alongside upper body motions. It has a "torso" that extends from a thick rolling platform (like a huge Roomba). Overall, the robot's aesthetic is somewhat chunky with smooth white limbs and joints, a cartoonish face and a plastic "ballcap." Despite its appearance and deliberate movements, Memo is surprisingly nimble and capable of completing basic household tasks. It can perform various activities in the kitchen and living areas, such as cleaning plates and glasses, loading them into the dishwasher, removing clothing, making espresso and even assisting in the preparation of simple dishes. Humans intuitively learn and understand how to interact with their environment because of the wiring of their own brains. As they become more comfortable within an environment, their dexterity improves. AI robots, on the other hand, depend on vision action models that don't have this grounding and must be trained in completing even basic tasks such as picking up objects, closing cupboards and navigating obstacles like tables and chairs. To overcome this challenge, Sunday developed what they call the Skill Capture Glove, a wearable that captures how people move, clean and organize in their own homes. Using a pair of these gloves, the company's researchers built a dataset of around 10 million genuine household routines captured in over 500 homes, representing a massive amount of data diversity. According to Sunday, the combined domestic training of this dataset enables Memo to adjust to the unpredictable nature of home environments, including kitchens, living rooms and laundry areas. "Tony's ALOHA and Cheng's UMI research told us that with enough data, dexterous manipulation tasks are actually possible with pretty low-cost hardware," said Camilla Guo, head of product at Sunday. ALOHA, or a low-cost open-source hardware system, developed at Stanford University in collaboration with Google LLC's DeepMind AI division, sought to build low-cost robotics platforms using imitation learning to train them to perform complex tasks. UMI, or universal manipulation interface, provided a data collection and policy learning framework for teaching robots how to handle tasks using human demonstrations, paving the way for the company's skill capture gloves. Sunday launched with a noteworthy $35 million in funding backed by Benchmark and Conviction. "The promise of AI robotics isn't back-flipping or dancing demos, but robots that work in messy, real-world situations," said Eric Vishria, general partner at Benchmark. "To have those, we need real-world training data. We have about one-millionth of the data we need." Sunday said applications for the company's Founding Family Beta program, which will put the robots in homes, opened on November 19. From the applicants, the company will select 50 households to become early adopters of Memo in late 2026. "There are plenty of companies building on our work to advance their research," said Zhao, "but what we are building here is something that is even larger: to put a robot in every home."