NewORA v0.1 is out: read the post

ORA

Open Robot-Use Agents

Open models operate robots. Zero-shot.

ORA is an open harness for robot use. The model looks, reasons and writes programs; a supervisor admits every motion; every run is sealed and can be replayed.

Get started Paper Code Gallery

One recorded run for each task ORA v0.1 solves: 27 of RoboDojo's 40 tasks and 90 of RoboLab's 100, each marked as it succeeds. DeepSeek-V4.1-Flash, code mode, zero-shot; sped up.

v0.1What the agent does now

Four changes since October's preview. Each opens its part of the post.

  1. Write code.The model writes Python against the same robot tools; one turn carries 3.6 robot actions instead of 1.1.
  2. Move smoothly.Every move is planned around what the depth cameras see; moves stopped by a collision fall from 14.1% to 7.0% on RoboDojo.
  3. Work across open models.In code mode on 39 RoboDojo tasks, six open models solve from 48.7% (DeepSeek) down to 5.1% (Nemotron).
  4. Cold-start post-training.On RoboLab-120, Qwen3.8-27B solves 5.4% alone, 52.5% inside ORA, and 61.9% with PEAR's training on top.

ResultsZero-shot, beside published systems

ORA in code mode with DeepSeek-V4.1-Flash, an open-weight model, on the cases of RPent's leaderboard.

RoboTwin C2R

ORA (ours)79.6
RPent + GPT-5.562.4
LingBot-VLA50.4
π0.547.9

LIBERO-Pro

ORA (ours)88.0
RPent + GPT-5.575.1
π-RLinf50.0
π0.511.0

ORA, zero-shotRPent: LLM + VLA + task memoryVLA trained on demonstrations

Rows other than ORA are RPent's published leaderboard (snapshot of 2026-09-22). ORA: 235 of RPent's 250 RoboTwin cases; LIBERO-Pro Task + Swap, initial states 0–4. RoboLab, RoboDojo and the same model in other systems are in the post.

MethodHow ORA works

The model asks for an action; the harness carries it out on the robot and returns what happened; the model decides again. A turn carries one tool call, or one program.

The agent loop, around a robot

Language modelplans, grounds, decides
HarnessManualrules, body card, task, tool manual
ToolsLookMeasureAct
Contextrecent turns in full, older turns as decisions and lessons
trust boundaryRobotbehind the trust boundary, with a success check the model never sees

Tool calls. Each turn the model sends one tool call; the harness runs it and one result comes back before the model decides again.

Code mode. Each turn the model writes one program, which calls the same tools from a persistent workspace, up to 12 robot motions admitted one at a time, before one result comes back.

What one turn carriestool callsprograms
Robot actions per acting turn1.103.62
Reasoning per action, characters33.4k12.7k
One loop, two interfaces. A turn carries one tool call or one program. Numbers: DeepSeek-V4.1-Flash, 59 RoboDojo cases in each setting.

Run it

One harness for RoboTwin, LIBERO-Pro, RoboDojo, RoboLab and a real arm, with any OpenAI-compatible or Anthropic model.

curl -fsSL https://raw.githubusercontent.com/fast-rua/ORA/preview/v0.1/scripts/install.sh | sh -s -- --ref preview/v0.1

Read the docsOpen recorded runs in the galleryAll posts

ORA in the blog

All posts