ABot-PhysWorld: Interactive World Foundation Model for Robotic Manipulation with Physics Alignment

arXiv 26.3 2026 71.6 method, application

TLDR

A 14B Diffusion Transformer video world model for robotic manipulation that ensures physical plausibility via DPO post-training and introduces a zero-shot benchmark.

Reasoning

The paper addresses a critical limitation of video world models (physical implausibility) with a novel DPO-based training framework and a new benchmark, achieving SOTA. Strengths include clear problem definition, strong empirical results, and a new evaluation protocol. Weaknesses are the domain specificity to manipulation and reliance on a curated dataset.

Read-first score

Read-first score 71.6, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 61.

Recency 8%

100

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

Topical relevance 42%

87.1

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 25%

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

Reproducibility 25%

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

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 140.

Keyword Scores

world model

video world model

generative world model

interactive world model

world dynamics prediction

world simulator

model-based reinforcement learning world model

Deep Analysis

Innovations

DPO-based post-training framework with decoupled discriminators to suppress unphysical behaviors while preserving visual quality
Parallel context block for precise spatial action injection enabling cross-embodiment control
EZSbench: first training-independent embodied zero-shot benchmark combining real and synthetic unseen robot-task-scene combinations with decoupled protocol for physical realism and action alignment

Methodology

ABot-PhysWorld is a 14B Diffusion Transformer model trained on a curated dataset of three million manipulation clips with physics-aware annotation. It uses a novel DPO-based post-training framework with decoupled discriminators to suppress unphysical behaviors, and a parallel context block for precise spatial action injection enabling cross-embodiment control. Evaluation is performed on PBench and the newly introduced EZSbench benchmark.

Key Results

ABot-PhysWorld achieves new state-of-the-art performance on PBench and EZSbench, surpassing Veo 3.1 and Sora v2 Pro in physical plausibility and trajectory consistency.

Resources

Paper