Robots learn physics by watching 1M hours of humans

Training robots has always had an unglamorous dirty secret: humans had to put on VR headsets or grab teleoperation rigs and babysit hardware for endless hours just to build a dataset.

Redwood City-based Dyna Robotics just flipped that script on August 10 by introducing DYNA-2. The foundation model skips robot-specific data entirely during pre-training, building its physical common sense solely by watching humans work.

Learning physics the way kids do

DYNA-2 is built as a World-Action Model trained on more than 1 million hours of first-person human video. That is roughly equivalent to 170 years of non-stop daily life recorded through human eyes.

The footage covers everyday dexterous tasks like folding laundry, cooking meals, cleaning surfaces, and assembling parts. Instead of simply cataloging what objects look like, the model predicts the next frame of action to understand how physical materials deform, balance, and respond when hands touch them.

The critical breakthrough is a scaling law: as the volume of human video data scales up, the model’s performance on real robot tasks climbs alongside it. Once pre-training is complete, adapting the brain to an entirely new robot chassis requires only a few hours of localized fine-tuning.

Dyna Robotics reports that this video co-training approach delivered a 133% boost in instruction-following tasks where robots execute physical movements from prompts.

From warehouse picks to hotel make-readies

Putting that brain into production, Dyna showcased its flagship physical agent, Dyna-2.1, targeting high-dexterity workflows across logistics, hospitality, and dining.

In warehouse fulfillment, the robot handles 200 units per hour across pick, inspect, place, and seal routines, maintaining plus-or-minus 2 mm placement accuracy even on novel items it has never seen before.

In hotel operations, a single unit turns over a checkout room to guest-ready status in 22 minutes, hitting over 97% on brand-standard inspection checklists without human intervention. In commercial kitchens, it scoops and plates 100 dishes an hour, clocking under 40 seconds per plate.

Crucially, it also features real-time recovery. Traditional factory arms freeze when something slips; this system senses slip-ups on the fly, adjusts hand position, and finishes the task without throwing an error code.

Why it matters

Data scarcity has been the single biggest bottleneck choking physical AI. Teleoperating humanoid arms for thousands of hours simply does not scale across the millions of manual tasks in the economy.

We think Dyna Robotics found a remarkably elegant workaround. By offloading spatial reasoning and physical intuition to the infinite supply of human first-person video, they shrink the hard robotic part down to a few hours of downstream fine-tuning. This could drastically lower the deployment barrier for factories, logistics hubs, and hotels eager to automate labor-short operations.

The catch is touch. Watching videos teaches visual causality, but it cannot convey grip force, tactile friction, or torque feedback. Fine-tuning will still need to bridge that tactile gap so fingers do not crush fragile cargo. Still, proving that robots can build a workable world model just by watching us move is a massive leap forward for the entire industry.