Are AI robots actually close to being any good?

A viral Reddit thread has sparked a major debate on the realistic capabilities of today's AI-powered robots, questioning whether they can perform the dynamic, unstructured physical work of skilled tradespeople. The original poster highlighted the stark contrast between polished laboratory demonstrations and the chaotic reality of a job site, using the example of a robot autonomously diagnosing and fixing a leak under a sink in an unfamiliar home. This frames a critical bottleneck in robotics: moving from scripted, controlled environments to generalizable physical intelligence.

Experts and commenters largely agree that while AI has made leaps in digital reasoning with models like GPT-4 and Claude 3, translating that to the physical world is exponentially harder. Current robots from companies like Boston Dynamics and Figure excel at specific, pre-programmed movements or teleoperated tasks but lack the integrated suite of capabilities—advanced computer vision for cluttered spaces, fine motor dexterity with tools, and common-sense reasoning about novel situations—required for true autonomy. The consensus is that AI may replace certain cognitive jobs long before it masters the embodied intelligence of a plumber or electrician.

The discussion underscores that the biggest hurdles are not just hardware but the AI 'brain.' Robots need to combine large language models for instruction understanding with vision-language-action models for perception and precise control, all operating reliably in safety-critical scenarios. This points to a future where robots may first assist tradespeople with specific sub-tasks rather than replace them entirely, with full autonomy likely being a decade or more away for such complex work.