AIAny
Icon for item

ZimaBlue: Evolving Generalizable World Action Models through Scalable Video Pre-training

Learns generalizable World Action Models for robotic manipulation by scaling causal egocentric video pretraining and grounding learned dynamics with heterogeneous robot trajectories. Key features: a three-stage curriculum (video pretraining, video-action mid-training with a unified action representation, and target-robot specialization) and a Slow–Fast dual-system for 30 Hz real-time action prediction.

Introduction

Robotic manipulation often fails to generalize because collecting diverse, action-labeled trajectories is costly; ZimaBlue flips that bottleneck by treating large-scale egocentric video as the primary source of embodied experience and showing how to convert it into executable control.

Key Findings
  • Scaling causal embodied video pretraining to over 120,000 hours yields large gains in real-robot zero-shot performance (reported improvement from 36.1% to 77.8%), suggesting broad video priors substantially improve out-of-distribution generalization. This implies pretraining on diverse human and robot first‑person video can substitute for—rather than merely supplement—target-robot data.
  • Three-stage curriculum: (1) causal video pretraining to learn spatiotemporal and physical regularities from egocentric videos; (2) video-action mid-training that grounds visuals into robot control using a unified action representation across embodiments; (3) target-robot post-training for embodiment-specific calibration. The separation helps each data source play its natural role: world knowledge from video, action grounding from mixed robot data, and precise control from target demonstrations.
  • System architecture: a Slow high-capacity world model for generalizable spatiotemporal representations paired with a lightweight Fast branch for 30 Hz action prediction on commodity GPUs, making generative WAMs viable for real-time control.
What Sets It Apart
  • Emphasis on causal, future-predictive video pretraining rather than relying solely on action-labeled robot data, enabling transfer from diverse human egocentric footage.
  • A unified action representation that standardizes states/actions across heterogeneous robots, reducing overfitting to a single embodiment and enabling cross‑embodiment transfer.
  • Practical deployment focus: asynchronous Slow–Fast dual-system balances representation quality and runtime action throughput.
Who It's For and Tradeoffs

Great fit if you need improved generalization for robot manipulation tasks and can leverage large-scale egocentric or mixed robot video corpora. Particularly relevant to researchers and engineers working on embodied AI, video-based pretraining, and multi-embodiment control. Look elsewhere if your target domain has strict safety/regulatory constraints that forbid relying on human-video priors without extensive verification, if you lack any heterogeneous robot data to perform mid-training grounding, or if low-latency on constrained hardware prohibits the Fast branch's runtime requirements.

Practical takeaway

ZimaBlue presents a scalable recipe: use vast egocentric video to learn world dynamics, connect those dynamics to control with mixed-embodiment trajectories via a unified action space, then fine-tune to the deployment robot. The approach trades more upstream data and a two-branch runtime design for substantially better zero-shot and out-of-distribution performance.

Information

  • Websitearxiv.org
  • OrganizationsJoy Future Academy
  • AuthorsXionghao Wu, Yijun Yang, Shiyang Zhou, Haoze Sun, Jianhui Liu, Songsong Yu, Jiyao Zhang, Wenbo Li, Bo Wang, Guoqing Ma
  • Published date2026/08/31

More Items

Develops a vision-language foundation model for autonomous driving that unifies 3D BEV perception, visual question answering, and motion planning without changing the pretrained VLM architecture. Key elements include an external BEV perception head for 3D detection and occupancy, a Planning Expert using flow-matching for trajectory prediction, and a staged training recipe combining driving and general VLM data.

Converts posed indoor RGB(-D) video into editable, simulation-ready 3D scene graphs by parsing multi-view evidence into per-object bundles, generating complete object assets from that evidence, and placing them with GizmoAct, a VLM policy that refines 9-DoF poses through closed-loop GUI actions.

Proposes GenFirst, a generation-before-reconstruction end-to-end training strategy for latent generative models that avoids latent collapse by prioritizing generative objectives and then progressively strengthening reconstruction, validated with strong gFID/GenEval results on ImageNet-256 and text-to-image tasks.