Architecture

VAM

2025ExperimentalPublished: 27 July 2026Updated: 27 July 2026Published

Key innovation

Uses a generative video model — which already understands world dynamics from large-scale pre-training — as the core of the robot policy, adding an action decoder instead of training a policy from scratch on static images (as in VLA models).

How it works

1) A base generative video model is pre-trained at scale on footage, learning the dynamics and behavior of the world. 2) For a given task and observation, the model predicts future frames/video sequences (how the scene should unfold). 3) An action decoder converts these predicted frames into robot motion commands (e.g. trajectories, manipulator control). 4) The model is fine-tuned to a specific task on a small amount of robot data (demonstrations), because most of the dynamics knowledge comes from video pre-training.

Problem solved

Vision-Language-Action models trained on static images require very large amounts of robot data and generalize poorly to new scene dynamics. VAM reduces data requirements and improves dynamics understanding by transferring knowledge from generative video models.

Components

Generative video modelPolicy core / dynamics prior

A pre-trained video-generation model that understands world dynamics and behavior and predicts future scene frames.

FLUX 3 (Black Forest Labs)The video model used in FLUX-mimic.

Official

Action decoderVideo -> action mapping

A module converting predicted frames/video sequences into robot motion commands (trajectories, manipulator control).

Official

Implementation

Reference implementations

mimic-video (mimic robotics)

mimic robotics

Official

FLUX-mimic (mimic robotics + Black Forest Labs)

mimic robotics

Official

Implementation pitfalls

Video-inference latencyHigh

Frame generation can be too slow for real-time control.

Fix:Distillation, shorter prediction horizon, inference optimization, dedicated hardware.

Video-action mismatchMedium

Predicted frames may be physically infeasible for the robot (the decoder must ground them).

Fix:Train the decoder on robot data, kinematic constraints, expert corrections.

Evolution

2025

mimic-video — the first Video-Action Model

Inflection point

mimic robotics introduces the Video-Action Model concept, combining video generation with robot action prediction.

2026

FLUX-mimic — a VAM built on FLUX 3

mimic robotics and Black Forest Labs present FLUX-mimic; task fine-tuning from ~30 min of robot data, deployed at Audi Production Lab.

Hyperparameters (configurable axes)

Video backboneCritical

Choice of the generative video model that forms the core.

FLUX 3Used in FLUX-mimic.

Action decoder typeHigh

How predicted frames are mapped to motion commands.

Fine-tuning data amountHigh

Amount of robot data needed to fine-tune a task.

~30 minvs 30+ h in prior approaches.

Computational complexity

Time complexity: Zalezna od bazowego modelu wideo. Space complexity: O(rozmiar modelu wideo + dekoder).

Compute bottleneck

Generative video-model inference

Predicting video frames/sequences at control time is the most compute-intensive part; achieving latency sufficient for real-time robot control is a key challenge.

Depends on

Rozdzielczosc i dlugosc horyzontu wideo