Understand this story
The English summary is being updated. See the reporting timeline below.
Latest development · Report
Hao AI Lab Releases ORA Traces, Featuring 3K+ Robot Task Trajectories from 7 ModelsDevelopments
1 confirmed developments
- Source:UCSD Hao AI Lab
Reports and discussion timeline
Browse reports and discussion by time. Their evidence levels remain distinct.
Nov 19
UCSD Hao AI LabFirst-partySelected
Hao AI Lab Releases ORA Traces, Featuring 3K+ Robot Task Trajectories from 7 Models
Hao AI Lab has released ORA Traces, featuring 3K+ task trajectories from 7 models across 5 benchmarks, along with system prompts, camera images, and complete CoT. https://t.co/QBts601SyK
UCSD Hao AI LabFirst-partySelected
RoboDojo Analysis Shows Models Write and Reuse Robot Manipulation Tools on the Fly
Hao AI Lab analyzed 34,327 functions written by models across 727 RoboDojo tasks and found that the models create tools on the fly based on the workspace. These include perception and geometry helper functions, which the models also re-import and use in subsequent steps.
UCSD Hao AI LabFirst-partySelected
Hao AI Lab Releases Zero-Shot Robotic Agent Framework ORA v0.1
Hao AI Lab has released ORA v0.1, an open-source zero-shot robotic agent framework. Using DeepSeek V4.1-Flash, it achieved a 39.9% completion rate across the same 198 RoboDojo cases, compared with 27.8% for Uni-Robot-Agent. Compared with the October preview, ORA improved from 5.2% to 39.9% on RoboDojo and from 71.5% to 79.6% on RoboTwin. The team also released ORA Traces, covering 7 models, 5 benchmarks, and 3K+ episodes, and launched Project FastRUA.
UCSD Hao AI LabFirst-partySelected
DeepSeek Completes 20 of 59 Paired Cases in RoboDojo Code Mode
Using the same robotics tools, DeepSeek completed 20 of 59 paired RoboDojo cases in code mode, compared with 6 in tool-calling mode. The statistics include cases completed after exceeding the step limit. Tool-calling mode stopped after 94 model calls; in code mode, the robot completed the tower-stacking task.
UCSD Hao AI LabFirst-party
UCSD Hao AI Lab Releases Zero-Shot Robotic Manipulation Framework ORA v0.1
UCSD Hao AI Lab has released ORA v0.1, an open zero-shot robotic manipulation framework that increases the RoboDojo completion rate on DeepSeek V4.1-Flash from 5.2% to 39.9% and the RoboTwin completion rate from 71.5% to 79.6%. ORA enables models to write tools on the fly and uses cuRobo to plan actions based on depth maps; the team also released ORA Traces, comprising 7 models, 5 benchmarks, and 3K+ episodes.
AI Trends