Instant re-inference and no interpolation lift a baseline policy from 76% to 88%

DominiqueCAPaul · x · 2026-08-04

The author says setting RTC to do instant re-inference and removing interpolated steps raised success rate from 76% to 88% on a baseline policy.
The improvement suggests that a more direct control loop can outperform a smoother but slower intermediate-step setup.
The tradeoff is that the resulting behavior looks very jerky, yet it still works surprisingly well.
With no further context in the post, the main takeaway is that this control-policy tweak materially improved task success despite uglier motion.

More from Embodied