Repolode
Submit
hao-ai-lab

hao-ai-lab/ORA

An open agent harness that lets open-weight language models operate robots zero-shot on LIBERO, RoboTwin and RoboDojo simulation tasks.

View on GitHub
Stars
112
Forks
4
Open issues
0
Last commit
17 days ago

What is it?

ORA is an agent harness for robot manipulation. It connects an open-weight language model, through an OpenAI-compatible or Anthropic API, to a robot or simulator, and lets the model call tools for perception, grasp proposals and motion planning. The project targets zero-shot operation: no vision-language-action model, no robot training and no task-specific memory are required.

The harness supplies the model with the task, a robot manual and tool contracts, and manages long tool results within an episode. A supervisor checks and dispatches every action the model proposes, so the model cannot bypass it. The repository includes three benchmark environments, LIBERO / LIBERO-Pro (Panda), RoboTwin 2.0 (dual-arm ALOHA) and RoboDojo (ARX on Isaac Sim 5.1), each running in its own isolated interpreter.

Every model request, tool call, action and receipt is recorded and can be replayed with a terminal UI or browsed in a web-based evidence viewer. Users can also send messages to a running agent, and such runs are marked as human-assisted.

Key features

  • 01Connects to OpenAI-compatible or Anthropic model APIs, hosted or self-hosted.
  • 02Provides perception (SAM3 segmentation, depth geometry), grasp proposals and CuRobo motion planning as model tools.
  • 03A supervisor checks and executes every action before it reaches the robot.
  • 04Runs LIBERO, RoboTwin 2.0 and RoboDojo benchmarks with their official success checks.
  • 05Records and seals every request, action and receipt, with replay in a terminal UI and a web viewer.

Use cases

  • Evaluating open-weight language models on robot manipulation benchmarks without training.
  • Comparing agent harness designs for embodied tasks in simulation.
  • Inspecting and replaying agent decisions, tool calls and actions from past runs.
  • Steering a running robot agent with messages from a terminal interface.

Quick start

  1. 1

    Clone the repository

    git clone https://github.com/hao-ai-lab/ORA.git && cd ORA
  2. 2

    Install ORA's locked Python environment

    uv sync --all-extras --locked
  3. 3

    Set your model API key

    export DEEPSEEK_LOW_API_KEY=sk-...
  4. 4

    Run a bounded check episode with no simulator or GPU

    uv run --locked ora run --env synthetic

Things to know

  • Robot benchmarks need a Linux machine with an NVIDIA GPU of 32 GB or more; the README verifies only specific GPUs.
  • Running the benchmarks requires a model API key, and the default example uses a DeepSeek endpoint.
  • A PyPI release is still pending, so installation currently comes from a source checkout.

In collections

Robotics & embodied AI projects

Topics

#robotics#llm-agent#embodied-ai#simulation#manipulation#benchmark#zero-shot

FAQ

What is ORA?
An open agent harness that lets open-weight language models operate robots zero-shot on LIBERO, RoboTwin and RoboDojo simulation tasks. ORA is an agent harness for robot manipulation. It connects an open-weight language model, through an OpenAI-compatible or Anthropic API, to a robot or simulator, and lets the model call tools for perception, grasp proposals and motion planning. The project targets zero-shot operation: no vision-language-action model, no robot training and no task-specific memory are required.
Is ORA free to use? What license does it have?
ORA is open source under the Apache-2.0 license. Check the license terms before using it in commercial products.
Is ORA still maintained?
Actively maintained. The last commit to ORA was on Oct 21, 2026.
What can I use ORA for?
Evaluating open-weight language models on robot manipulation benchmarks without training. Comparing agent harness designs for embodied tasks in simulation. Inspecting and replaying agent decisions, tool calls and actions from past runs. Steering a running robot agent with messages from a terminal interface.
How do I install ORA?
Clone the repository: git clone https://github.com/hao-ai-lab/ORA.git && cd ORA Install ORA's locked Python environment: uv sync --all-extras --locked Set your model API key: export DEEPSEEK_LOW_API_KEY=sk-... Run a bounded check episode with no simulator or GPU: uv run --locked ora run --env synthetic
Where can I get ORA?
The source code and installation instructions are on GitHub: https://github.com/hao-ai-lab/ORA

Summary generated by AI from the project's README. GitHub stats refresh daily.

Related projects