WorldFoundry
English中文

LingBot-World

Method overview: Pipeline of LingBot-World. The left part shows the pipeline of LingBot-World video generation. LingBot-World uses an image or a video, noisy latents, and user-defined action signals as inputs to generate video.

LingBot-World from Robbyant (Ant Group): world and camera control; Verified; dedicated `lingbot-world`.

Edit this pageGitHubMarkdown

LingBot-World is Robbyant's camera-controlled world video model. In this catalog it is fully verified: the base-cam checkpoint completed full-quality artifact validation on 2026-07-28 using the official 161-frame, 70-step camera-control demo — including a four-A100 compact topology that adapted FSDP/Ulysses settings and produced a coherent artifact.

Official paper teaser: Interactive world simulation across diverse environments. The figure showcases selected samples generated by LingBot-World , demonstrating its capability to synthesize high-fidelity videos in various domains,.
Official paper teaser: Interactive world simulation across diverse environments. The figure showcases selected samples generated by LingBot-World , demonstrating its capability to synthesize high-fidelity videos in various domains,.
Method overview: Pipeline of LingBot-World. The left part shows the pipeline of LingBot-World video generation. LingBot-World uses an image or a video, noisy latents, and user-defined action signals as inputs to generate video.
Method overview: Pipeline of LingBot-World. The left part shows the pipeline of LingBot-World video generation. LingBot-World uses an image or a video, noisy latents, and user-defined action signals as inputs to generate video.

The recorded paper is arXiv 2601.20540.

Project page: technology.robbyant.com/lingbot-world. Code: robbyant/lingbot-world.

Build your run

Pipeline
LingBotPipelineworldfoundry.pipelines.lingbot_world.pipeline_lingbot_world
Task
Environmentlingbot-world
DeviceCUDA 12.1
Runnerworldfoundry.pipeline
Loadingfrom_pretrained
Invocationnative_or_standard_invocation
worldfoundry-eval run \
  lingbot-world \
  --pipeline.task-profile image-navigation-video \
  --json

What it is for

lingbot-world is the recorded camera-control route. Start here to follow a catalog camera path.

The official base-cam recipe uses eight GPUs with DiT/T5 FSDP plus Ulysses sequence parallelism; Workspace also supports a four-GPU compact topology (both 4 and 8 divide the model's 40 attention heads). A shared timestep embedding preserves floating precision for integer scheduler timesteps, preventing distributed sequence-parallel output corruption. It is published by Robbyant (Ant Group).

Catalog tasks include world, camera control, and world model.

Recorded evaluation entry points include CameraBench (Camera-control adherence of world rollouts), WorldModelBench (General world-rollout quality).

LingBot world + camera control. Video / VLA / map are sibling cards.

Runtime contract

Optional image, prompt, interactions, action_path, and frames; artifacts lingbot-world.mp4.

CLI task profiles: image-navigation-video and image-action-video. WorldFoundry binds LingBotPipeline.

Family status: Verified.

Recorded image-navigation-video defaults: frames=161, seed=42 Recorded image-action-video defaults: frames=21, seed=42.

The wrapper resolves only local HFD or explicit checkpoint directories — no runtime downloads. Base-cam runs in the dedicated lingbot-world environment; Base-Act is separately cataloged. The official Fast checkpoint uses Base-Cam's shared T5/VAE assets plus its own 16 shards. The public binding produced a fully decodable 21-frame, 544 × 720, 16 fps Fast clip with changing content on one H100. Select it with runtime_variant=fast and fast_model_path; Fast frame counts must be 9, 21, 33, … because its three-latent-frame chunks otherwise truncate the clip. Longer Fast rollouts and camera adherence remain unmeasured.

Running in WorldFoundry

Launch worldfoundry-eval run lingbot-world --pipeline.task-profile image-navigation-video to follow the camera path. The command builder above switches --pipeline.task-profile among recorded tasks and still emits worldfoundry-eval run.

Recorded weight repositories include robbyant/lingbot-world-base-cam. Prepare the environment and check local assets with bash scripts/inference/prepare_model_infer.sh lingbot-world. Add --download to fetch missing public checkpoints.

Status: Verified.