UCSD Hao AI Lab Releases Zero-Shot Robotic Manipulation Framework ORA v0.1
UCSD Hao AI Lab has released the open-source zero-shot robotic manipulation framework ORA v0.1. DeepSeek V4.1-Flash achieved 39.9% on RoboDojo and 79.6% on RoboTwin. The team tested six open models using the same prompts and tools, finding the strongest correlation with Terminal-Bench. It also released ORA Traces, which contains more than 3K complete trajectories from seven models across five benchmarks.
Why read this
SelectedAI score 66
Experiments using standardized prompts and tools connect robotic manipulation with terminal benchmarks, providing a transferable basis for model selection and agent training comparisons.
Topics
Body
Which models are good at robot use? We ran six open models with the same prompt and tools and compared the result with public benchmarks. We found t lines up best with Terminal-Bench, where a model works in a terminal. https://t.co/8H5ida0Dsk
AI Trends