OpenDriveLab/UniVLA

Health v2 · maintenance only

Flags: no_releases

Adoption not part of the score

1134 stars · 68 forks observed · 2026-09-24

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

UniVLA is an open-source framework for training cross-embodiment vision-language-action (VLA) robot policies using task-centric latent actions extracted from videos. It provides the full training recipe—latent action model, generalist policy pretraining, and post-training/evaluation pipelines for benchmarks like LIBERO, CALVIN, and real robots.

Use cases

• train a vision-language-action policy for robot manipulation
• learn robot policies from cross-embodiment and human videos
• evaluate VLA models on LIBERO and CALVIN benchmarks
• deploy a generalist policy on a real robot arm
• extract task-centric latent actions from videos
• pretrain a robot foundation model with less compute than OpenVLA

When to choose

• you need a compute-efficient, state-of-the-art VLA training recipe
• you want to leverage heterogeneous video data including human demonstrations
• you need cross-embodiment transfer for manipulation or navigation tasks

When to avoid

• you need a production-ready robot control stack rather than research code
• you lack GPU resources for large-scale model training
• your task requires a single-embodiment, plug-and-play controller with no training

Facets

library · maturity active

machine-learning deep-learning llm-training simulation robotics machine-learning autonomous-vehicles python vla vision-language-action robot-learning cross-embodiment latent-actions manipulation research-code linux gpu

Member repositories

For agents

markdown · JSON · MCP: product_card(name="OpenDriveLab/UniVLA")