Skip to content
fast-ruaPublic

About

ORA: Open Robot-Use Agents

Resources

Contributing

Security policy

Stars

128 stars

Watchers

0 watching

Forks

Repository files navigation

ORA: Open Robot-Use Agents. Open models operate robots. Zero-shot.

ORA is an open harness that lets a language model operate a robot through camera images and tools:
open-weight models, zero-shot, with no VLA, no robot training and no task memory.

Latest release License: Apache-2.0 Documentation Blog post: ORA v0.1

Website · Blog · Docs · Gallery · Quick start · Results · How it works · Citation

ORA v0.1 teaser: RoboDojo zero-shot results, a model writing a program that builds a tower, solved episodes on four benchmarks, and what v0.1 adds
From the v0.1 film. Watch all 91 seconds, with sound, or open the runs in the gallery.

New in v0.1

v0.1 follows October's preview. With the open-weight DeepSeek-V4.1-Flash it solves 39.9% of RoboDojo zero-shot. Read the post.

A program on a screen and the hopping path it drives
Write code
The model writes Python against the same robot tools; one turn carries 3.6 robot actions instead of 1.1.
A smooth path through three objects beside a jagged one
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.
Several model chips plugged into one harness that drives a gripper
Work across open models
In code mode on 39 RoboDojo tasks, six open models solve from 48.7% (DeepSeek) down to 5.1% (Nemotron).
A small model lifting a barbell
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.

Release notes · Changelog · 2026/11/18 ORA v0.1 · 2026/10/15 ORA preview

Results

DeepSeek-V4.1-Flash, zero-shot, each benchmark's own success check, every success counted. The runs behind these numbers open in the gallery.

Benchmark Robot Cases Code mode Tool calls
LIBERO-Pro Franka Panda 800 87.5% 85.4%
RoboTwin 2.0, clean to randomized two-arm ALOHA 235 79.6% 80.4%
RoboLab, 100 of its 120 tasks DROID Franka 500 79.2%
RoboDojo two-arm ARX X5 198 39.9%

RoboLab and RoboDojo limit an episode's simulator steps; within the limit ORA solves 67.0% of RoboLab and 10.6% of RoboDojo. Baselines on the same cases and every count: how successes are counted, the numbers in full.

Quick start

1. Install ORA and a simulator. On a Linux machine with one NVIDIA GPU of 32 GB or more; no sudo.

curl -fsSL https://raw.githubusercontent.com/fast-rua/ORA/main/scripts/install.sh | sh
ora setup --env robodojo

2. Run a task. Store your model key once (DeepSeek's API by default), then run:

ora providers key DEEPSEEK_LOW_API_KEY
ora run --env robodojo --task general_pickup --confirm-launch

The terminal shows the model's words, its tool calls and the robot's cameras, and the run is sealed when it ends. No GPU at hand? The first episode runs the same loop on a synthetic robot.

Documentation

To Read
install by hand, or see what the installer does Installation
use OpenAI, Anthropic or a self-hosted model Model APIs
pick a benchmark: LIBERO-Pro, RoboTwin 2.0, RoboDojo, RoboCasa365, RoboLab Supported setups · One-GPU guide
watch, replay or export a run (ora watch, ora web, ora replay) Run and inspect
give the robot instructions yourself (ora play) Free play
run a real robot (I2RT YAM station) Real robot cells
run a benchmark at scale Official suites
look up a command or fix a failure Command reference · Troubleshooting

How it works

The agent loop around a robot: a language model sends a call or a program to the harness, with its manual, tools and context; the harness runs it on the robot behind a trust boundary and returns what happened

An agent works in a loop. The model asks for an action; the harness, with its manual, tools and context, carries it out on the robot and returns what happened; the model decides again. In code mode one turn carries a program instead of a single call. Every motion passes a supervisor, the benchmark's own check gives the verdict, and every request, action and receipt is sealed so a run can be replayed. How ORA works · Architecture

Versions

main is the latest release, and every release keeps its tag: v0.1.0 is this one, and v0.1.0-preview is October's preview, unchanged. To stay on a version, check out its tag and follow that version's README; from v0.1.0 on, install.sh | sh -s -- --ref TAG installs a given release. What changed and what an upgrade needs is in the changelog.

Contributing

Start contributing covers the first hour and CONTRIBUTING.md the path to a pull request. To report a bug, open an issue; for a security problem see SECURITY.md.

Acknowledgements

We thank the RPent team for releasing their evaluation cases, leaderboard and code (Harness VLA); the authors of CaP-X and ASPIRE, whose work inspired ORA's tools; and DeepSeek for releasing DeepSeek-V4.1-Flash as an open-weight model. We thank Zhaoning Wang for deploying ORA on a real robot, and Runyu Lu and Nan Xue for helpful discussions.

ORA builds on open-source work: LIBERO, LIBERO-Pro, RoboTwin 2.0, RoboDojo, RoboCasa365 and RoboLab as benchmarks; SAM 3, Contact-GraspNet and cuRobo as tools; and pi-tui for the terminal UI. Code adapted from RoboTwin and CaP-X keeps its original license; see NOTICE.

Citation

@misc{ora2026v01,
  title        = {{ORA} v0.1: Robot-Use Agents Now Write Code, Move Smoothly, and Reach 40\% on {RoboDojo}},
  author       = {Yuanbo Yang and Junda Chen and Quanquan Peng and Yulun Wu and Shaoxiong Duan and Hao Zhang},
  year         = {2026},
  month        = nov,
  howpublished = {FastRUA Blog},
  url          = {https://fastrua.ai/blog/2026-11-code-mode/}
}

Licensed under Apache-2.0. Bundled third-party code retains the notices in NOTICE.

About

ORA: Open Robot-Use Agents

Resources

Contributing

Security policy

Stars

128 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages