LAC paper: shifting RL architecture burden to the critic cuts robot inference latency 4x

heghbalz · x · 2026-09-05

A new offline RL paper questions why actors bloat with slow diffusion/flow-matching models while the critic is discarded after training.

In real-world robotics, multiple denoising steps per decision destroy control frequency
LAC shifts the burden to training: a residual MLP backbone, n-step bootstrap targets, and categorical cross-entropy stabilize a deep critic against optimization collapse and bootstrap drift
With the deep critic doing the heavy lifting, a bare-bones deterministic actor matches heavy generative baselines on OGBench
Single forward pass yields up to 4x lower inference latency

More from Embodied