---
type: analysis
slug: revisited-five-weeks-on-world-action-models-are-faster-but-edge-control-remains-unproven
title: "Five weeks on, world-action models are faster—but edge control remains unproven"
description: "New work strengthens the case for combining visual prediction and action generation, with reported speed and benchmark results. But those results are on datacenter-class hardware, not onboard robot computers, so the original analysis’s central edge-deployment implication remains a hypothesis."
published_at: 2026-10-31T13:00:00.000Z
topics: ["trends","retrospective"]
source_urls: ["https://newruntime.com/analysis/fast-world-action-models-will-push-robot-control-onto-the-edge/","https://arxiv.org/abs/2610.13740","https://github.com/iitimii/SmolWAM","https://arxiv.org/abs/2610.06617"]
author: "Andrey Reshetnikov"
author_profile: https://newruntime.com/owner-profile.md
---
**Verdict: partly held.** The efficiency and single-model direction gained support, but the evidence still does not show that these models can deliver reliable, low-latency control on robot edge hardware.

Original analysis: [Fast world‑action models will push robot control onto the edge](https://newruntime.com/analysis/fast-world-action-models-will-push-robot-control-onto-the-edge/), published 2026-09-26 (5 weeks ago).

## What we said

We argued that BFL’s FLUX 3 Action showed how a single-step model generating both video and robot actions could make low-latency manipulation more practical, and that step-distillation could reduce inference cost. We urged builders to test the open weights and LoRA recipe on their own embodiments, while noting that the evidence was vendor-reported and limited to simulation. Our strongest forward-looking implication was that the approach could enable on-device control and hybrid planning, potentially simplifying the stack.

## What happened since

- A subsequent world-action model, SmolWAM, reported predicting actions and future visual latents together with 0.88 billion parameters. Its authors reported 95.5% success over 400 LIBERO episodes and 57 milliseconds per 16-action chunk on an NVIDIA A100. That is a concrete result in the same general direction—joint world and action prediction with strong benchmark performance—but the measured latency is on an A100, not a robot’s onboard computer. [1]
- SmolWAM’s code and weights were released, and its project says inference runs on CUDA GPUs, Apple silicon, and CPUs. Its published comparison reports 11 times lower latency and 18.5 times less compute than its teacher on an A100. This makes experimentation and access across hardware types more plausible, but portability is not evidence that an embedded system can meet a real-time control budget. [2]
- RealtimeWAM paired one-step action generation with asynchronous video/action inference and reported near-lossless benchmark performance alongside up to roughly 25 times end-to-end speedup on an H100. This supports the idea that inference bottlenecks can be reduced without abandoning world-action models. However, the result is also a datacenter-GPU benchmark and does not establish edge deployment or lower total system cost. [3]

## What we missed

The original analysis moved too quickly from fast single-step inference to the prospect of on-device manipulation. The new evidence is encouraging about model efficiency and benchmark performance, but all the cited speed results are on A100 or H100 hardware; even the reported ability to run on CPUs and Apple silicon does not establish useful performance in a robot’s onboard loop. We also framed the opportunity as a potential stack simplification without evidence here that a joint model reduces real-world system costs or safely replaces separate planning and control modules. The original caveat about simulation and robustness still matters, and these findings do not close it.

## Forecast check

- BFL will announce an early‑access partnership with at least one robotics platform (e.g., a warehouse robot vendor) that provides a sandbox for fine‑tuning FLUX 3 Action on physical hardware. (55%, due 2027-03-26): **unclear**. The forecast is not due until 2027-03-26. The available items describe research and releases, not an early-access physical-robotics partnership or hardware sandbox.
- The RoboLab‑120 leaderboard will receive at least three new entries that cite FLUX 3 Action as a baseline before the competition deadline in 12 months. (45%, due 2027-09-26): **unclear**. The forecast is not due until 2027-09-26. The available items do not report new RoboLab-120 leaderboard entries citing FLUX 3 Action.

## What to do now

Builders should treat world-action models as promising candidates for experiments, not as evidence that edge control is ready to deploy. If evaluating one, measure end-to-end latency and task success on the actual onboard hardware and robot embodiment, including perception, action execution, and any asynchronous components; do not extrapolate A100 or H100 timings to an embedded device. Compare against the existing policy and planning stack on safety, robustness, and total system cost, not just inference speed or simulated success. The original pilot idea remains useful as a test plan, but its suggested sub-100-millisecond target should be measured rather than assumed achievable, and a successful simulation or desktop run is not a deployment result.

## Next check

The decisive next observation is an independently reported physical-robot trial showing end-to-end control latency and task reliability on the robot’s onboard hardware.

## Sources

- [Fast world‑action models will push robot control onto the edge](https://newruntime.com/analysis/fast-world-action-models-will-push-robot-control-onto-the-edge/) (original analysis, 2026-09-26)
1. [SmolWAM: A Latent World-Action Model for Robotic Manipulation on Edge Devices](https://arxiv.org/abs/2610.13740) (2026-10-31)
2. [SmolWAM GitHub repository](https://github.com/iitimii/SmolWAM) (2026-10-31)
3. [RealtimeWAM: One-Step Asynchronous World Action Models](https://arxiv.org/abs/2610.06617) (2026-10-31)

Author: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- collection: [Analysis](https://newruntime.com/analysis.json): Open the latest authorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)