open-gigaai/giga-world-policy

Health v2 · maintenance only

Flags: no_license

Adoption not part of the score

1398 stars · 107 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-30, confidence not recorded

GigaWorld-Policy is a World Action Model (WAM) for robot policy learning that jointly models actions and future visual observations during training while using action-only decoding at inference. Its 0.5 release uses a Mixture-of-Transformers architecture to achieve 85ms inference latency on a local RTX 4090 for real-time closed-loop robot control.

Use cases

• train robot policies with world models
• real-time robot control with low latency inference
• learn robot manipulation policies from visual demonstrations
• run a world action model locally on a single GPU
• research action-conditioned world modeling for robotics

When to choose

• you need efficient real-time robot policy inference on consumer GPUs
• you want to leverage future visual dynamics as dense supervision for policy learning
• you are researching world models or action models for robot control

When to avoid

• you need a general-purpose vision-language model rather than a robot control policy
• you lack GPU hardware for training or inference
• you need a plug-and-play robot stack without ML research involvement

Facets

library · maturity active

machine-learning deep-learning llm-training simulation robotics machine-learning deep-learning artificial-intelligence python world-model robot-policy-learning action-model mixture-of-transformers real-time-inference robotics pytorch gpu linux

Member repositories

For agents

markdown · JSON · MCP: product_card(name="open-gigaai/giga-world-policy")