Agent skill · research science · aradotso
nano-world-model
Minimalist batteries-included repository for training, evaluating, and deploying diffusion-forcing video world models for robot manipulation, gaming, and MPC planning.
Why this skill is useful
Adds executable scripts for training, evaluating, and deploying video world models that are not commonly found in public documentation.
What it needs
Requires conda, git installed locally. About 5k tokens when loaded. Last updated 2026-07-09. 70 stars on the source repository.
What this skill does
Nano World Model Skill by ara.so — Daily 2026 Skills collection. Nano World Model is a minimalist, batteries-included repository for training video world models using diffusion-forcing transformers. It supports diverse domains (robot manipulation, gaming, simulation), long-horizon autoregressive rollouts, video-to-3D reconstruction, and MPC-style planning via CEM. --- Installation Download the I3D model used for FID/FVD evaluation: --- Environment Variables / Path Configuration Set these before running any command, or define them in src/configs/local/paths.yaml: Or create src/configs/local/paths.yaml: --- Key CLI Commands The main entry point is src/main.py with Hydra config composition. Training Evaluation Long-Horizon Rollout MPC Planning (CEM) Video → 3D Point Cloud --- Pretrained Checkpoints Load from HuggingFace directly: Available checkpoints: Domain HF Repo Steps -------- --------- ------- Point Maze knightnemo/nanowm-b2-dino-wm-point-maze-30k 30k Wall knightnemo/nanowm-b2-dino-wm-wall-15k 15k Rope knightnemo/nanowm-b2-dino-wm-rope-15k 15k Granular knightnemo/nanowm-b2-dino-wm-granular-15k 15k PushT knightnemo/nanowm-b2-dino-wm-pusht-100k 100k RT-1 knightnemo/nanowm-b2-rt1-300k 300k CSGO knightnemo/nanowm-l2-csgo-100k 100k Use with CLI: --- Configuration System (Hydra) Config is composed from src/configs/. …
How to use it
Reference it in AdaL, Claude Code, Cursor or any coding agent — nothing to install:
@skills aradotso/nano-world-model