Benchmark few-step action generators against JEPA Policy

Conduct a like-for-like comparison between JEPA Policy and few-step action generators, including distilled and single-stage alternatives, under matched evaluation conditions.

Background

JEPA Policy is evaluated against MIP and a 100-step Diffusion Policy configuration, while the paper explicitly notes that it does not provide a matched comparison with few-step action-generation methods. Such a comparison would clarify whether the reported performance and latency advantages persist against more computationally efficient diffusion, consistency, or shortcut-based policies rather than only against the evaluated baselines.

The unresolved comparison is especially relevant because the paper characterizes its own predictor as a two-pass, diffusion-free policy and states that few-step alternatives already exist. A controlled benchmark would therefore distinguish gains attributable to paired future-representation supervision from gains attributable primarily to the inference budget or sampling strategy.