Learned latent dynamics roll forward from few observations to recover hopper state with uncertainty

hisspikeness · x · 2026-10-08

Final part of a research thread (7/): in ongoing work, the learned latent dynamics can be rolled out from a few initial observations (t<0) to recover the MuJoCo hopper's signals, including uncertainty estimates. An earlier tweet in the thread (6/) shows the contrast: adding a pixel-reconstruction loss to the same model makes the latents partly encode nuisance factors (colors, lighting, camera), so the state can no longer be decoded. Together this demonstrates the representational power of latent predictive modeling under nuisance and noise.

Original post →

Learned latent dynamics roll forward from few observations to recover hopper state with uncertainty

More from Embodied