TLDR Tech / Practical
ORA: Open Agents Control Robots Without Prior Training
ORA, an open robot-use agent framework, allows open-weight language models to operate robots in a zero-shot manner. This means no visual-language-audio (VLA) models, robot training, or task-specific memory are needed. ORA has demonstrated competitive performance on RoboTwin and LIBERO-Pro benchmarks, achieving 69.2% and 81.4% success rates respectively, matching agents that utilize frontier models.
Read the full story