Adithya Balaji
Integration view

Paradigm comparison

VLA vs world action model vs task-specific diffusion policy vs RL policy vs classical pipeline. In real systems these are layers of one stack, not rivals: a classical safety and control layer at the bottom, RL-trained controllers for physics-dominated skills, a VLA or WAM for visuomotor behaviour, and a reasoner on top.

··

Where world action models sit

Taxonomy after the 2026 WAM surveys: cascaded models predict a future and then extract actions; joint models predict both together. Futures can be pixels (video) or latents (features). Inference cost falls from top-left to bottom-right.

World action model taxonomyTwo-by-two map: cascaded versus joint architectures, pixel versus latent future representations, with example models in each cell.FUTURE REPRESENTED AS PIXELS / VIDEOFUTURE AS LATENTS / FEATURESCASCADEDpredict the future,then extract actions· UniPi, Gen2Act· π0.7 generated subgoal images· NovaFlow, Dream2Flow (flow → action)highest fidelity, highest latency· VPP (video-model features → policy)· Latent action planners (GO-1)· Latent plans for VLA headscheaper inference, less visual foresightJOINTpredict futures andactions in one model· DreamZero (14B video diffusion)· Cosmos Policy (actions as latent frames)· X-WAM, Fast-WAM, LingBot-VAhighest fidelity, highest latency· LaWAM (foundation-feature futures)· V-JEPA 2-AC, VLA-JEPA· GigaWorld-Policy (video only in training)cheaper inference, less visual foresightReference point: a VLA has no explicit future prediction; it maps observations straight to actions through a VLM backbone.

Sources: WAM survey (Sep 2026), awesome-wam-2026 taxonomy. Detail in the frontier entry.

··

How the paradigms stack in one robot

LayerTypical paradigm (2026)RateWhy this paradigm
Reasoning / planningVLM / embodied-reasoning model (ER 2, Cosmos Reason 2), sometimes cloud0.1–2 HzSemantics, long horizon, language; latency tolerant
Visuomotor policyVLA today; WAM or WAM-distilled policy emerging5–15 Hz chunksGeneralist manipulation from pixels and language
Skill residualsRL residual or small diffusion policy20–50 HzPrecision and speed on contact phases
Whole-body / balance (humanoids)RL motion-tracking controller50 Hz (→1 kHz PD)Physics-dominated; trained in sim
Control & limitsClassical impedance, IK, safety filters0.5–1 kHzGuarantees and certification
Safety functionsClassical, safety-rated hardwarems cycleRequired by standards; independent of learning