EgoGenesis: Egocentric World-Action Modeling with Online Anchored Projective Memory and Action-3D RoPE

Source

• Source name: arXiv
• Source kind: paper
• Authors / organization: Zexuan Yan, Yuzhou Wu, Yue Ma, Zonghang He, Kaibo Yin, Xiaobing Tu, Yinggui Wang, Jinkui Ren, Xiantao Zhang, Shijian Wang, Jinghong Liu, Linfeng Zhang
• Venue: arXiv: cs.CV
• Published: 2026-07-30
• Collected: 2026-07-31
• URL: http://arxiv.org/abs/2607.28243v1
• PDF: https://arxiv.org/pdf/2607.28243v1
• DOI: UNKNOWN

Plain-Language Brief

The authors present a method called EgoGenesis for creating synthetic manipulation videos for embodied AI using egocentric video. The method uses a pretrained video generation prior and introduces two geometry-aware conditioning mechanisms: Online Anchored Projective Memory (OAPM) and Action-3D Rotary Position Embedding (A3D-RoPE). These components improve the visual fidelity, geometric stability, and action alignment in long egocentric rollouts. When 400 real trajectories are augmented with 400 synthesized trajectories from EgoGenesis, it leads to improved success rates on single-arm and dual-arm tasks for a downstream WAM (a type of robot arm).

What the Document Says

• The authors present a method called EgoGenesis.
• EgoGenesis is an egocentric world-action simulator that synthesizes controllable, high-quality manipulation videos to expand scarce real-world training data.
• EgoGenesis builds on a pretrained video generation prior and introduces two geometry-aware conditioning mechanisms: Online Anchored Projective Memory (OAPM) and Action-3D Rotary Position Embedding (A3D-RoPE).
• OAPM preserves a first-frame 3D scene anchor while periodically refreshing a recent state during autoregressive generation.
• A3D-RoPE encodes end-effector motion with camera-aware 3D rotary coordinates, injecting action geometry into skeleton-to-video cross-attention for precise control.
• These components improve visual fidelity, geometric stability, and action alignment in long egocentric rollouts.

Key Concepts Mentioned

• EgoGenesis
• egocentric video
• manipulation videos
• synthetic data
• online anchored projective memory (OAPM)
• action-3D rotary position embedding (A3D-RoPE)

Claims Made by the Document

• EgoGenesis improves visual fidelity, geometric stability, and action alignment in long egocentric rollouts.
• Augmenting 400 real trajectories with 400 EgoGenesis-generated trajectories improves out-of-distribution real-robot success rates.

Evidence Used

• Improved visual fidelity, geometric stability, and action alignment in long egocentric rollouts is stated as a result of the components introduced by EgoGenesis.
• The improvement in out-of-distribution real-robot success rates is demonstrated through the comparison between 400 real trajectories and 400 augmented trajectories.

Methods / Approach

• Pretrained video generation prior
• Online Anchored Projective Memory (OAPM)
• Action-3D Rotary Position Embedding (A3D-RoPE)

Limitations Stated by the Authors

The authors do not explicitly mention any limitations of their method.

Limitations Noted by the Agent

Agent interpretation — not from the source.

• The quality and generalizability of the synthetic data generated by EgoGenesis may be dependent on the quality and characteristics of the pretrained video generation prior.
• The success rates mentioned in the text are based on single-arm and dual-arm tasks, and it is unknown how well EgoGenesis would perform on other types of tasks.

Open Questions

• What is the pretrained video generation prior used by EgoGenesis?
• How does the success rate of EgoGenesis compare to other methods for creating synthetic manipulation videos?

Notes for Future Q&A

May include interpretation. Labeled where it does.

The paper does not provide details on the specifics of the pretrained video generation prior or how it was trained.

Raw Abstract / Excerpt

Egocentric video offers rich manipulation experience for embodied AI, yet collecting diverse egocentric data across scenes, objects, motions, and embodiments remains costly. We present \method, an egocentric world-action simulator that synthesizes controllable, high-quality manipulation videos to expand scarce real-world training data. \method{} builds on a pretrained video generation prior and introduces two geometry-aware conditioning mechanisms. Online Anchored Projective Memory (OAPM) preserves a first-frame 3D scene anchor while periodically refreshing a recent state during autoregressive generation. Action-3D Rotary Position Embedding (A3D-RoPE) encodes end-effector motion with camera-aware 3D rotary coordinates, injecting action geometry into skeleton-to-video cross-attention for precise control. Together, these components improve visual fidelity, geometric stability, and action alignment in long egocentric rollouts. Moreover, augmenting 400 real trajectories with 400 \method-generated trajectories improves out-of-distribution real-robot success from 77% to 84% on single-arm tasks and from 53% to 70% on dual-arm tasks, demonstrating that the synthesized data substantially improve downstream WAM generalization.