@jchen-eth · PhD Candidate · Zürich, CH
Evaluating whether scaling episodes actually beats scaling model size.
Key claim: action chunking with a frozen vision backbone matters more than parameter count below 10M episodes. I reproduced the door subset on real hardware and got within 3 points.
16 is our sweet spot for 20 Hz control. 32 gets smoother trajectories but recovers badly from perturbations.
Environment diversity contributes roughly 3x more to out-of-distribution success than raw episode count past 5K per task.