SC3-Eval: Evaluating Robot Foundation Models via Self-Consistent Video Generation

arXiv 2026 59.5 method

TLDR

SC3-Eval adapts a video foundation model into a policy evaluator using three consistency forms to simulate robot rollouts without real-world costs.

Reasoning

The paper introduces a novel self-consistent video generation method for evaluating robot policies, addressing key challenges like compounding errors and multi-view consistency. Its strengths lie in the three complementary consistency mechanisms, but the abstract lacks quantitative results or comparisons to baselines, limiting assessment of empirical effectiveness.

Read-first score

Read-first score 59.5, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 48.

Recency 6%

100

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

Citation impact 18%

91.5

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. citation_normalized_percentile=0.91487073

Topical relevance 29%

68.6

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

Methodology quality 18%

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

Reproducibility 18%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=reproduce

Citation velocity 12%

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

Field roles

FoundationFrontierBridge

Rank sensitivity

Stability: volatile; rank range: 425.

Keyword Scores

video world model

world model

world dynamics prediction

world simulator

generative world model

interactive world model

model-based reinforcement learning world model

Deep Analysis

Innovations

Forward-inverse dynamics consistency that jointly trains the model to predict frames from actions and recover actions from frames, anchoring rollouts to a physically plausible action manifold.
Cross-view consistency that trains the model to inpaint each camera view from the other, maintaining multi-camera coherence over long rollouts without explicit memory.
Test-time consistency that reuses the inverse dynamics mode as an uncertainty signal to terminate rollouts when generated frames drift from requested actions.

Methodology

SC3-Eval adapts a pre-trained video foundation model into a policy evaluator by enforcing three forms of consistency. Forward-inverse dynamics consistency jointly trains the model to predict frames from actions and recover actions from frames. Cross-view consistency trains the model to inpaint each camera view from the other. Test-time consistency uses the inverse dynamics mode at inference as a per-action-chunk uncertainty signal to terminate drifting rollouts.

Key Results

Across seven real-world vision-language-action policies, SC3-Eval achieves a closed-loop Pearson correlation of 0.929 and MMRV of 0.119, outperforming three strong prior video-model-based baselines, and generalizes to new tasks while reproducing failure modes seen in real-world rollouts.

Resources

Paper