Open Source RoboticsRobots, software & learning

LingBot-World

Robotics model

Open-weights interactive video world simulator built on Wan2.2 for camera- and action-conditioned rollout generation (16 FPS, minute-level horizon in v1; Apache-2.0 for v1 code and checkpoints, while separate v2 causal release is CC BY-NC-SA 4.0). [1] [2] [4] [5]

Setup & usage →CheckpointsModel hub / homepageGetting started

For
Embodied AI and world-model researchers; Robot learning simulation teams; Multi-GPU inference engineers [2] [4]
Parameters
18,544,332,864 (~18.5B F32 parameters in lingbot-world-fast safetensors metadata [2] [3] [4]
Embodiments
Embodiment-agnostic egocentric / free-camera visual world simulator (conditioned on camera trajectories and action… [2] [4]
Source rights
Named licenses are documented by part. Includes noncommercial terms. Review licenses

Source researched · Not tested by us
Sources checked · Research details

Licenses & related hardware

Requirements · Research details and dates

On this page

Overview

LingBot-World is an open-weights interactive video world simulator built on the Wan2.2 i2v-A14B image-to-video Diffusion Transformer architecture (paired with a T5 text encoder and video VAE). Rather than outputting physical robot joint commands, it generates future video frames conditioned on 6-DoF camera pose trajectories or discrete action strings (--action_path, --action_string, --allow_act2cam). [1] [2] [4]

Model card summary

Architecture

Wan2.2-based image-to-video Diffusion Transformer (i2v-A14B task architecture with T5 text encoder and video VAE) conditioned on camera poses or discrete action strings for interactive long-horizon world simulation.

[2] [4]

Base model

Wan2.2 (Wan-Video/Wan2.2 i2v-A14B architecture)

[4]

Parameter count

18,544,332,864 (~18.5B F32 parameters in lingbot-world-fast safetensors metadata; i2v-A14B MoE/DiT architecture; base-cam repo total 149.25 GB includes dual/unquantized components)

Exact tensor count for lingbot-world-fast verified via Hugging Face API safetensors metadata.

[2] [3] [4]

Input modalities
rgblanguage

Initial RGB image, text prompt, and camera pose trajectories or action strings (e.g., WASD / camera control; action-string conditioning noted in README).

[2] [4]

Output modalities
video-frames

Generates 480P (480x832) and 720P (720x1280) interactive video rollouts at 16 FPS (up to 961 frames / ~1 minute).

[2] [4]

Action space

Camera pose trajectories (6-DoF camera extrinsics/intrinsics) and user-friendly discrete action strings (--action_path / --action_string / --allow_act2cam) conditioning future video frame generation rather than low-level robot joint torques.

[4]

Control frequency

16 Hz (16 FPS video generation reported in model card)

Model card reports sub-1-second latency at 16 FPS for real-time interactive rollout.

[2]

Trained / supported embodiments
  • Embodiment-agnostic egocentric / free-camera visual world simulator (conditioned on camera trajectories and action strings)

[2] [4]

Training datasets
  • Undisclosed / non-public pretraining video corpus (arXiv:2601.20540); camera poses estimated via ViPE in examples

Training datasets are not publicly released for download in the repository.

[4] · Editorial assessment

Interfaces
  • PyTorch CLI (generate.py / generate_fast.py)
  • Camera trajectory / action-string conditioning
  • Hugging Face safetensors

[2] [4]

Checkpoints

CheckpointFormatSizeAccessPinned revision
robbyant/lingbot-world-base-cam (v1, Apache-2.0)safetensors149.25 GBopen6fc824ffc338
robbyant/lingbot-world-fast (v1 distilled, Apache-2.0)safetensors74.18 GBopenb7a1db2da9bd
robbyant/lingbot-world-v2-14b-causal-fast (v2, CC BY-NC-SA 4.0 noncommercial)safetensors80.16 GBopen5c33dd40b213

robbyant/lingbot-world-fast reports 74,182,934,990 bytes (~74.18 decimal GB / 69.09 GiB) usedStorage in S003 and requires downloading base-cam first per S004.

[2] [3] [5]

Is this for you?

Use cases
  • Interactive camera- and action-conditioned rollout generation
  • Long-horizon visual world simulation for robot learning
  • Real-time fast rollout generation via distilled LingBot-World-Fast

[2] [4]

Known limitations
  • Heavy multi-GPU compute requirement (8 GPUs with FSDP + Ulysses parallelism documented for default 480P/720P inference; 149.25 GB base-cam / 69.09 GB fast checkpoint footprint)
  • Training datasets and training code are not released in the inference repository
  • License split across generations: v1 (robbyant/lingbot-world and v1 checkpoints) is Apache-2.0, whereas v2 (robbyant/lingbot-world-v2 and v2 checkpoints) is strictly noncommercial CC BY-NC-SA 4.0

[2] [3] [4] [5]

Safety and limitations
  • Outputs simulated video frames rather than closed-loop physical robot motor commands; editorial adoption caveat: supplied evidence does not establish physical accuracy.
  • Strict version-dependent licensing and maintenance tradeoff: v1 code and weights are Apache-2.0 but v1 maintenance is discontinued (S004), while v2 code and weights are noncommercial CC BY-NC-SA 4.0.
  • Training code and pretraining datasets are not publicly released; minimum per-GPU VRAM is unspecified (8-GPU torchrun command is an upstream reference workflow, not a verified minimum).

[1] [2] [5]

Compute & fine-tuning

Inference compute
  • Documented hardware: 8x multi-GPU node (torchrun --nproc_per_node=8 with FSDP + Ulysses)
  • Precision: bf16
  • Reported latency: 1000 ms
  • Edge / consumer capable: No

Model card reports latency under 1 second (<1000 ms) when producing 16 FPS; exact minimum VRAM per GPU not stated.

[2] [4]

Fine-tuning
  • Supported: No

Public repository provides inference/rollout generation scripts only; training/fine-tuning scripts are not included.

[4] · Editorial assessment

Host computer

Documented reference commands use 8 GPUs (torchrun --nproc_per_node=8, --ulysses_size 8, FSDP); community 4-bit quantized variant mentioned in README for limited VRAM.

[4]

Operating system
  • Linux

Inferred from documented Linux/CUDA Python environment and shell instructions.

[4] · Editorial assessment

Dependencies
  • Python
  • PyTorch
  • Wan2.2 codebase dependencies
  • Diffusers
  • T5 / VAE assets

[2] [4]

Skills
  • Python
  • PyTorch Distributed (torchrun / FSDP)
  • Diffusers / Wan2.2 inference

Inferred from documented Python/PyTorch/CLI setup and training dependencies.

[4] · Editorial assessment

Setup & usage

Open the upstream getting-started guide [4] · Model card / documentation [1] [2]

  • Clone robbyant/lingbot-world and install Wan2.2 dependencies
  • Download robbyant/lingbot-world-base-cam or robbyant/lingbot-world-fast from Hugging Face
  • Run torchrun --nproc_per_node=8 generate.py (or generate_fast.py) with --dit_fsdp --t5_fsdp --ulysses_size 8

[4]

Upstream-reported benchmarks

Upstream-reported only: Open Source Robotics has not run training, inference, or hardware evaluations for these benchmarks.

BenchmarkUpstream-reported scoreScope & caveats
LingBot-World technical report evaluations (arXiv:2601.20540)Upstream model card and README give 961 frames (~1 minute at 16 FPS) as an example setting conditional on sufficient CUDA memory, and report sub-1s latency.Upstream-reported world-model metrics; not independently verified by OSR.

[2] [4]

Project notes

Visual simulation scope and multi-GPU compute requirements

LingBot-World is an embodiment-agnostic visual rollout simulator, not a closed-loop physical robot controller; as an editorial adoption caveat, the supplied evidence does not establish physical accuracy for generated video rollouts. Checkpoint storage requirements are large and additive for the distilled variant: robbyant/lingbot-world-fast reports 74,182,934,990 bytes of repository storage (~74.18 decimal GB / 69.09 GiB; 18,544,332,864 F32 parameters in safetensors metadata), and S004 instructs Fast users to download robbyant/lingbot-world-base-cam first and place the Fast weights inside that directory—so Fast's individual storage does not represent total setup storage or GPU memory. Upstream's documented inference workflow uses an 8-GPU node (torchrun --nproc_per_node=8 with FSDP and Ulysses parallelism) for 480P and 720P generation, with community 4-bit quantization noted separately; minimum per-GPU VRAM is unspecified in the supplied documentation, and 8 GPUs describes upstream reference commands rather than a verified minimum configuration. [2] [3] [4]

Upstream-reported world-model metrics

The model card and README describe an example setting of 961 frames (~1 minute at 16 FPS) conditional on sufficient CUDA memory, alongside sub-1s latency claims (linking to arXiv:2601.20540, which is not excerpted here). All reported horizon, frame-rate, and latency figures are upstream-reported; Open Source Robotics has not run inference or benchmarked LingBot-World on multi-GPU hardware. [2] [4]

What we checked

None. Prepared from public upstream sources without hands-on hardware execution.

Updates

Upstream last updated
Jul 9, 2026 [1]
Latest release
Apr 2, 2026 [3]
Profile updated
Nov 21, 2026
Sources last checked
Nov 21, 2026
Next review
May 21, 2027
Reviewed revision
robbyant/lingbot-world@a43bec7f8091c83e9b30b16b912f6fc906236fa6; HF lingbot-world-base-cam@6fc824ffc338d64c97c77e2eb8c0f4cfc24d82bd; HF lingbot-world-fast@b7a1db2da9bd3ba675ee01c228f41db0309b981e · software repository

Separate license matrix

Major 2026 open-weights interactive world simulator with Apache-2.0 v1 inference code and weights; marked include_partial because training code and pretraining datasets are not released and v2 switches to noncommercial CC BY-NC-SA 4.0.

LayerAvailability / license
Codeavailable · Apache-2.0 (v1 inference repo); CC BY-NC-SA 4.0 (separate v2 repo)
Verified directly in LICENSE.txt of both repositories: robbyant/lingbot-world is Apache-2.0; robbyant/lingbot-world-v2 is CC BY-NC-SA 4.0. Note: inference code only (no training code). [1] [5]
Model weightsavailable · Apache-2.0 (v1 base-cam & fast checkpoints); CC BY-NC-SA 4.0 (v2 checkpoints)
Corrects preliminary audit note in candidates.json: v1 HF checkpoints (lingbot-world-base-cam, lingbot-world-fast) are Apache-2.0; only v2 checkpoints (e.g., lingbot-world-v2-14b-causal-fast) are CC BY-NC-SA 4.0. [2] [3] [5]
Training datasetsunavailable · License not fully assessed
Pretraining video datasets are not publicly released; only small inference example assets are included in the repo. [4]
Documentationavailable · Apache-2.0 (v1 README)
[1] [4]

Strict version split between v1 and v2 licenses and maintenance

Licensing and maintenance split strictly by version. The v1 repository (robbyant/lingbot-world) and v1 Hugging Face checkpoints (lingbot-world-base-cam and lingbot-world-fast) are licensed under Apache-2.0, verified in LICENSE.txt and Hub metadata. However, the reviewed v1 README (S004) explicitly states that robbyant/lingbot-world is no longer actively maintained and directs future development to LingBot-World-Infinity / v2 (robbyant/lingbot-world-v2 and robbyant/lingbot-world-v2-14b-causal-fast), which are licensed under noncommercial CC BY-NC-SA 4.0. Choosing permissive v1 weights therefore also means accepting discontinued v1 maintenance. In the reviewed v1 materials, only inference code and small example assets are published; training code and pretraining video datasets are not released there, and v2 training-artifact availability was not verified in this review. [1] [2] [3] [4] [5]

References

Each source links to the exact revision we read. “Used for” shows where this page relies on it.

Correction guidance · Discussion is not enabled · Profile format 0.8.0-draft.1

Change history

2026-11-21 — Initial 0.8.0-draft.1 model draft prepared from pinned GitHub commits a43bec7f8091 (v1) & 1895d300d8ac (v2) and HF revisions 6fc824ffc338 / b7a1db2da9bd / 5c33dd40b213.