In Karur, an Indian city, Objectways employees wear cameras on their heads and repeat hundreds of precise hand movements for long hours, folding towels without wrinkles. Each film is then broken down into prime factors: finger contours are marked, objects are tagged, gestures are classified, and faulty shots – with a missed movement or a towel placed incorrectly – are thrown out of the pool. This painstaking data crafting builds the context the algorithms need. It teaches them the sequence of actions, orientation in space and appropriate pressure. Objectways provides these described recordings to companies working on robotics and generative AI around the world.
On the West Coast, this data rush is described by Ulrik Stig Hansen, co-founder of Encord, a platform that manages large-scale annotation projects. In his opinion, as reported by Tech Spot, Robotics is experiencing a renaissance as companies race to build models designed for the physical world. Encord cooperates with players such as Physical Intelligence and Dyna Robotics, and the demo data it collects includes: from Objectways, become raw material for training models to perform multi-tasking, human activities.
Complete activities to be replaced by robots
This is the fundamental difference from chatbots. Large language models mimic speech and reasoning after reading online text, however models for robots must learn from a first-person perspective. They need an image from the camera “in the eyes” of the operator and information about the sequence, strength and trajectory of movements.
Google DeepMind research on the RT-2 and RT-X family of models has shown that combining web data with robotic demonstrations improves generalization and allows you to translate language commands into actions in reality. It's a step from reading the Internet to, for example, working in the kitchen.
At the same time, teleoperation is flourishing. Increasingly, robots – from arms to humanoids – are controlled remotely by people, sometimes even from another continent. Data from such sessions is streamed, analyzed, and the best fragments are used to supervise or train imitative models. “Hand farms” appear on the industry map – halls in Central and Eastern Europe, where rows of joysticks and operators spend hours teaching the machines new tricks. Critics respond that many humanoids under remote control look impressive, but when the “human leash” is removed, they still get lost in reality. For now.
Read also: You can now argue and dance with robots. Here are my impressions from the key technology fair
The scale is growing rapidly
Figure AI announced a partnership with Brookfield, one of the largest real estate owners in the world, to create an online-scale dataset of apartments, offices and commercial spaces. The goal is to teach the Helix model locomotion and manipulation in spaces designed for humansand the company has already started pilots in Brookfield environments.
High-profile previous demos, such as “teaching” a robot to make coffee after watching recordings, will now be provided with a solid base of data from real homes and corridors.
Scale AI, a company backed by Meta, has already raised over PLN 100,000. hours of analogous recordings for robotics in his laboratory in San Francisco. In this logic, companies also pay people in different countries to wear smart glasses and record ordinary activitiesbecause it is from them – from thousands of movements that we perform thoughtlessly – that “work maps” for future assistants are created. This is another version of the so-called the annotation economy that previously pushed autonomous cars forward.
Check also: Robots with AI will conquer the world. You can invest now [GOTOWE PORTFELE]
AI infrastructure giants are joining the race. Nvidia announced the GR00T project and the Isaac platform — hardware and software to accelerate the creation of generalized models for humanoids learning from language, video, and demonstrations. In 2025, the company announced the availability of GR00T N1 and additional elements of the ecosystem for generating synthetic data and physics simulations, strengthening the trend in which data from the real world is combined with data from the virtual world. This is an attempt to build an entire assembly line for physical AI.
Goldman Sachs estimates that by 2035 the humanoids market will reach a value of approximately USD 38 billion, and in the horizon of 2050 – according to UBS forecasts – it may grow to as much as USD 1.4-1.7 trillion, mainly due to the response to the demographic crisis and labor shortages. In investor scenarios, humanoids do not so much add excess productivity as plug structural gaps in services and care.
