Tech • AI • Robotics • Game
Robots are entering our homes… but can we trust them?
7/10TL;DR
Figure says its new Helix 2.5 AI sharply improves humanoid robots’ ability to complete household tasks in unfamiliar homes, highlighting fast progress but also major safety, labor and privacy concerns.
KEY POINTS
A jump in performance
Figure tested its Figure 03 humanoid in about 30 homes in the San Francisco Bay Area that had not been mapped or prepared in advance. Robots were asked to tidy 13 to 15 toys, fold towels and fully make a bed using the same objects across homes. With earlier control methods, full-task success was only 9%; with the new training system, branded Index, success rose to 56%.
Training from human video
The company’s approach relies on building a large visual library of everyday human actions. People equipped with cameras record tasks such as folding laundry, tidying rooms and handling objects, giving the system examples of hand position, grip and sequencing. The method mirrors the data-heavy training strategy that helped large language models improve rapidly.
A possible scaling law for robotics
The most important implication is not the current 56% score but the suggestion that more data and more computing power steadily reduce error rates. That resembles the scaling pattern seen in modern AI, where model performance improves predictably as training inputs grow. If that trend holds, domestic and industrial robotics could advance much faster than earlier generations that depended on painstakingly scripted motions.
From scripted demos to adaptive behavior
Earlier robot demonstrations often relied on highly controlled environments and preprogrammed action sequences. Newer systems increasingly adjust on the fly, including self-correction when a robot is badly positioned for a task. That marks a shift from choreography to more general physical reasoning, with future gains expected from richer world models designed for real environments rather than text alone.
Industry likely comes before the home
Household demos attract attention, but the clearest near-term market is industrial labor. A factory owner may see immediate value in a humanoid costing roughly $30,000 if it can replace repetitive work, while consumers may hesitate to bring a large autonomous machine into the home. In China, humanoids are already being discussed as a response to labor shortages and demographic decline, while Japan has focused more on elder care assistance.
China and the US are racing on different fronts
The current robotics race appears increasingly split between China and the United States. The US remains strong in frontier AI models, while China is pushing hard on manufacturing, deployment and data collection in factories. Public attitudes also differ sharply: enthusiasm for robots and AI is described as much higher in East Asia than in countries such as France, where skepticism remains widespread.
Safety remains a major barrier
A robot that succeeds only about half the time is still far from acceptable in a family home. Beyond simple failure, humanoids create physical risk: powerful motors, heavy limbs and rigid components can injure people if contact goes wrong. Some companies are already adding soft external materials to reduce impact, but safe operation around children, pets and cluttered rooms remains unresolved.
Alignment is harder in the physical world
The challenge is not just movement but judgment. A household robot must know when not to complete a command if the action creates danger, such as shutting a door on a child or mishandling a fragile object. In physical settings, bad decisions can cause immediate harm, making guardrails and common-sense reasoning more urgent than in purely digital AI systems.
A huge data market is emerging
Unlike text AI, robotics does not have a ready-made internet-scale archive of physical know-how. Companies are therefore trying to buy, license or collect vast amounts of embodied data. One estimate cited a current pool of about 500,000 hours of data versus a need closer to 100 million hours, suggesting a large new market for recorded manual skills and task demonstrations.
Work, know-how and ownership are at stake
The spread of humanoids raises direct questions about labor replacement, especially in factories and routine service jobs. It also touches intellectual property: rare manual techniques in sectors such as luxury goods, crafts and surgery could be digitized, copied, licensed or lost. The debate is no longer only whether robots can perform human tasks, but who controls the skills that train them and who benefits from the productivity they create.
CONCLUSION
Humanoid robotics is moving from spectacle toward practical deployment, with data-driven training now producing measurable gains in unfamiliar real-world settings. The next phase will depend not only on technical progress, but on whether companies and governments can solve safety, trust and social acceptance faster than the machines improve.
Ask a question