ZimaBlue lifts zero-shot robot manipulation from 36.1% to 77.8%

Scaling training from a single target robot's own demonstrations to over 120,000 hours of egocentric video pushed real-robot zero-shot manipulation success from 36.1% to 77.8%.

That result comes from ZimaBlue, a framework its authors describe as learning "generalizable World Action Models (WAMs) from large-scale video." The abstract's diagnosis is flat: "action-labeled robot trajectories are expensive to collect and inherently limited in diversity," while egocentric video offers "a far more scalable source of embodied experience."

Training runs in three stages. First, causal embodied video pre-training on human and robot egocentric footage. Then video-action mid-training that grounds the model in heterogeneous robot trajectories with "a unified action representation." Finally, specialization to a target robot for deployment. To serve control loops, the team pairs a "high-capacity Slow world model" with a "lightweight Fast branch enables 30 Hz action prediction on NVIDIA RTX 4090."

The paper reports "particularly pronounced gains on unseen tasks." The abstract publishes no per-task numbers and names no baseline comparisons.