Robot Foundation Models Hit 90% Success on 1M Video Hours

A Data Milestone for Embodied AI

A new wave of robotics announcements is putting hard numbers behind the promise of "foundation models" for physical machines. Dyna Robotics says it trained a new robot model on roughly one million hours of human video footage, and reports the resulting system hit a 90% success rate across a range of physical tasks 1. The scale of that training set marks a notable jump for embodied AI, an approach that borrows the same logic as large language models — feed a system enormous amounts of data and let it generalize — but applies it to bodies that must grasp, walk, and manipulate objects in the real world rather than just generate text.

Foundation Models Go Commercial

Dyna's announcement lands alongside other signs that robotics firms are racing to plug large-scale AI models into hardware rather than build everything in-house. Agile Robots has become the latest company to partner with Google DeepMind, agreeing to integrate DeepMind's robotics foundation models into its machines while simultaneously feeding data back to DeepMind's research teams 2. That arrangement reflects a broader industry pattern: robotics startups increasingly trade access to their hardware and real-world operating data for a share of the capability boost that comes from foundation models trained on vastly more information than any single company could collect alone.

Humanoids Head Toward the Battlefield

While Dyna and Agile Robots are chasing general-purpose task performance, other companies are pushing humanoid robotics into far more contentious territory. Foundation Future Industries, also referred to as Foundation Robotics Labs, is developing armed humanoid robots intended for U.S. military use, and has already secured roughly $24 million in Pentagon-linked contracts 3. The company has drawn scrutiny not only for the ethics of arming autonomous machines but also for its connections to the Trump family, with Eric Trump reportedly involved in the venture 36. Coverage notes that Foundation's 2024 launch broke with a long industry norm in which many robotics companies had pledged not to weaponize their platforms or let machines select targets without human authorization 5. Executives at the company have signaled an aggressive timeline, aiming to field military humanoid robots within 12 to 18 months 6.

The Bigger Picture: Jobs and Oversight

The rapid advance of physically capable, foundation-model-driven robots is arriving as broader anxieties mount over AI's effect on entry-level employment, with commentators warning that automation is reshaping hiring expectations faster than educational institutions can adapt 4. Taken together, the developments illustrate two converging trends: robotics companies are achieving measurable gains in real-world task competence through massive video and data training, while simultaneously courting political and defense-sector partnerships that raise fresh questions about accountability, labor displacement, and the militarization of humanoid machines.