Hydra-0: Action Flow for Generalist World Modeling and Control

arXiv 2026 44.6 method, application

TLDR

Hydra-0 represents robot actions as pixel motion, enabling a generalist world model for control and evaluation across embodiments and tasks.

Reasoning

The paper's core contribution is a novel action-flow interface for world modeling, supported by empirical results on the RoboLab benchmark and motion-error reductions. Weaknesses include limited visibility into methodology and potential overreliance on benchmark-specific metrics.

Read-first score

Read-first score 44.6, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 46.

Recency 6%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Methodology quality 18%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=baseline,benchmark,evaluation,result

Topical relevance 29%

65.7

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Reproducibility 18%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=none

Citation impact 18%

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. cited_by_count=0

Citation velocity 12%

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=0.00

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 384.

Keyword Scores

world model

interactive world model

world dynamics prediction

generative world model

world simulator

video world model

model-based reinforcement learning world model

Deep Analysis

Innovations

Introduces action flow, representing robot actions as pixel motion, as a shared visual interface for world modeling and control.
Enables generalist world modeling across embodiments, tasks, environments, and video-generation backbones by learning action consequences.
Supports zero-shot composition and data-efficient adaptation.
Demonstrates an emergent inverse mode that predicts compatible robot motion from desired object flow transferred from human demonstrations, with a trained action head mapping latent features to executable actions without task-specific expert robot demonstrations.

Methodology

Hydra-0 is a generalist world model conditioned on action flow, where robot actions are represented as pixel motion. It is evaluated against an action-conditioned baseline using robot-motion and object-motion error, zero-shot composition, data-efficient adaptation, and the RoboLab benchmark via Pearson correlation between replayed and reference success rates. An action head is trained on latent features from the inverse mode to produce executable actions from desired object flow.

Key Results

Hydra-0 achieves 90.4% lower robot-motion error and 60.2% lower object-motion error than the action-conditioned baseline, while supporting zero-shot composition and data-efficient adaptation. On RoboLab, it reaches a Pearson correlation of r=0.96 between replayed and reference success rates, and its inverse mode maps human-demonstration object flow to executable robot actions without task-specific expert demonstrations.

Resources

Paper