Robot AI is improving, but the humanoid home helper is not round the corner

Robots can now use vision-language-action models to handle tasks such as packing a lunchbox, but they are still likely to fail at jobs beyond their training examples. MIT Technology Review’s account explains why progress in robot AI is real, while claims of near-term general-purpose humanoids remain a much bigger leap.

Watch Desk analysis

What happened

The article describes a shift from hand-coded robot instructions to models that interpret scenes and produce movement commands. Vision-language-action models learn from demonstrations, often collected by people remotely operating robot arms. Google DeepMind’s Gemini Robotics, for example, has been demonstrated using ALOHA 2 arms to assemble a simple lunchbox and perform other trained tasks.

The catch is coverage: unlike language models, robotics has no vast pool of high-quality physical demonstrations to draw on. Human-operated data is costly to collect, video of people doing tasks can be poor training material, and real-world robot deployments bring reliability and safety challenges. Researchers are also exploring world models trained on video, 3D scans and sensor data to help robots predict what actions will do.

Key findings

• Demonstrations are making robots more capable
  Models can interpret a scene and act on learned examples, rather than relying only on extensive hand-coded rules.
• Novel tasks remain a weak spot
  A robot may fail when asked to do something outside its training set; today’s progress is not general competence.
• More data is not a settled answer
  Researchers disagree about whether collecting enough demonstrations can solve the problem, or whether different approaches are needed.

Why it matters

The gap between a robot that performs a rehearsed task and one that reliably copes with unfamiliar kitchens, objects and mishaps is the gap between a convincing video and a useful product. The article argues that humanoid appearance and general-purpose ability are often bundled together in public predictions, though one does not establish the other.

That makes the debate about methods as important as the launch promises. More demonstrations could broaden what robots can do, while world models might help them anticipate physical outcomes. Neither route, the article makes clear, has yet delivered a machine that can confidently handle the endless variations of everyday work.

Our read

This is a better story than either “robots are useless” or “your future housekeeper is in the post”. The evidence points to meaningful progress on trained tasks, alongside a stubborn generalisation problem. Enjoy the clever lunchbox demonstration; keep the moving date for the humanoid housekeeper in pencil.

What to watch

• Whether robot systems become more reliable on tasks they were not explicitly trained to perform.
• Whether world models improve performance in real environments, not just simulations.
• Whether companies publish evidence of useful, repeatable deployments beyond staged demonstrations.

Discussion spark: Should robotics companies prioritise improving performance on unfamiliar tasks, or focus on making a narrower set of jobs reliably useful first?

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.