One policy for the whole body changes the data problem
Splitting locomotion from manipulation was a good engineering decision that created a gap. Closing the gap with a single policy is the right fix, and it demands exactly the training data the field has least of.
DeepMind extended Gemini Robotics on 30 July with its first model controlling a humanoid's legs, torso, arms and hands under a single policy. That is a different engineering proposition from a locomotion stack bolted to a manipulation stack.
Why the split existed
Locomotion is a fast, high-frequency balance problem. Manipulation is a slower precision problem. Separating them let each be attacked with appropriate methods, and that was sound engineering rather than laziness.
The cost was everything falling between the two controllers: bracing against a surface to apply force, using body weight for leverage, catching balance while carrying something. Those are not exotic edge cases. They are most of what physical work actually consists of, which is why demos have leaned so heavily on tasks a stationary arm could do.
The fix relocates the bottleneck
A single policy closes the gap in principle and opens a data problem in practice. Whole-body control needs a training distribution covering coordinated whole-body behaviour — precisely what the field has least of. Teleoperation collects manipulation well because a human operator with a controller maps naturally onto arms and hands. It collects coordinated dynamic movement poorly, because a human cannot teleoperate balance.
Which is why every model in this category differs in architecture and agrees entirely on the constraint. Gemini Robotics, NVIDIA GR00T, Physical Intelligence, Figure Helix and Skild Brain all name too little robot data as the binding limit, with answers split between mass teleoperation, shared datasets like Open X-Embodiment, and synthetic data from world models.
Which makes the production numbers relevant
Figure 03 past 1,000 units, AgiBot at 15,000 cumulative, Atlas advancing, Optimus Gen 3 in low-volume ramp. The order-of-magnitude gap is two strategies, not a scoreboard: fewer capable units in demanding tasks, versus volume — and volume is what generates the operational data everyone just agreed is the bottleneck.
Unit counts remain the least informative metric available, which is exactly why they get quoted. A robot shipped is not a robot working. Deployment duration, task breadth and intervention rate would tell you far more, exist internally, and essentially never appear in public.
The real signal is that these are production figures at all. Two years ago the vocabulary was prototypes and demonstrations. Ramping manufacturing means supply chains, service networks and warranty exposure — expensive to commit to and awkward to reverse, which makes them more credible than any demo video.
BigGo Finance — Google unveils Gemini Robotics 2, tackling the whole-body intelligence challenge → · AgiBot — AgiBot reaches 10,000 units → · AI2Work — AgiBot ships 10,000 humanoid robots as China dominates global production →