Skip to content

Releases

ORA v0.1.0: robot-use agents now write code, move smoothly, and reach 40% on RoboDojo

Choose a tag to compare

@freemty freemty released this 18 Nov 23:49

ORA v0.1: RoboDojo zero-shot, a model writing a program that builds a tower, solved episodes on four benchmarks

ORA is an open harness that lets a language model operate a robot through camera images and tools, with no robot
training. v0.1 follows October's preview. With the open-weight DeepSeek-V4.1-Flash it solves 39.9% of RoboDojo
zero-shot. The post has the full account and the film with sound.

Highlights

  • Write code. In code mode the model writes a short Python program each turn that looks, measures, drives
    several motions and checks its own result before it hands control back. Programs run in a bubblewrap sandbox with
    no network and can call only the model's own tools; every motion still passes the supervisor. One turn carries 3.6
    robot actions instead of 1.1. Code mode is RoboDojo's default.
  • Move smoothly. Inverse kinematics, collision-aware planning (cuRobo in a world built from the depth cameras)
    and path timing inside joint limits (TOPP-RA and Ruckig) are one motion base shared by all five simulators. Moves
    stopped by a collision fall from 14.1% to 7.0% on RoboDojo. Where moves are planned on the policy side, a dry run
    shows the planned path as a camera-projected picture before the move runs, and a refused move returns the same
    picture.
  • Work across open models. Six open-weight models run the same harness unchanged. In code mode on 39 RoboDojo
    tasks they solve from 48.7% (DeepSeek-V4.1-Flash) down to 5.1% (Nemotron).
  • 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.

Also new

  • Free play. ora play loads a scene with no task. You give instructions one at a time and can talk to the
    agent while the robot moves: Enter joins the running instruction, Alt+Enter queues the next one, Esc stops.
    ora play --code-mode runs free play in code mode. Scenes exist for Synthetic, LIBERO, RoboTwin, RoboDojo and the
    YAM cell; play sessions are sealed and never scored.
  • The Pi agent core. The agent loop, request assembly, context management and provider calls run on the pinned
    Pi packages (TypeScript, ora-pi/); tools, robot semantics and authority stay in Python. It brings a coding
    agent's terminal: /model to switch models, /resume, /fork and /tree, your own skills and AGENTS.md. The
    Python Native core is retired.
  • A web page for every run. ora web RUN shows the conversation, each tool call or program, and every camera in
    the browser; ora replay RUN replays a finished run on the same page and --export out.mp4 makes a video.
    ora cameras draws the cameras beside the terminal.
  • Five benchmarks. RoboCasa365 (the 50 target tasks, PandaOmron mobile manipulator) and RoboLab (120 tasks,
    DROID Franka in Isaac Sim 5.1) join LIBERO / LIBERO-Pro, RoboTwin 2.0 and RoboDojo. Each runs in its own
    interpreter with the benchmark's own success check.
  • Evidence you can replay. Sealed runs carry a replay program; ora replay --check re-executes the recorded
    actions. Token use is counted exactly from the provider's usage.
  • Install in one line. curl -fsSL https://raw.githubusercontent.com/fast-rua/ORA/main/scripts/install.sh | sh
    installs uv and Node.js 22 when they are missing, checks ORA out and puts an ora command on your PATH.
    ora setup --env <env> then installs a simulator; for RoboDojo it also installs git-lfs and bubblewrap when the
    machine has none, and ora doctor reads the same settings. On a machine with several GPUs, ora run --gpu N runs
    on GPU N.

Results

DeepSeek-V4.1-Flash, zero-shot, each benchmark's own success check, every success counted.

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 the simulator steps of an episode. ORA records the benchmark's verdict at the limit and
lets the episode go on; within the limit it solves 67.0% of RoboLab and 10.6% of RoboDojo, and RoboDojo's leaderboard
counts only those. Baselines run on the same cases and every count are in the post's appendix:
how successes are counted and
the numbers in full.

Upgrading from the preview

  • Node.js 22.19 or newer is required: the Pi core is the only agent core. ora doctor checks it.
  • Store provider keys once with ora providers key NAME; they live in ~/.config/ora/pi-auth.json (mode 0600),
    never in a run.
  • The documentation site is built with Blume (./scripts/docs serve); MkDocs is gone.
  • Removed: ora-tui/ (the terminal UI is part of the Pi core), experiments/, ops/ and ci/.

Known limitations

  • RoboDojo: ORA reads the measured gripper opening where RoboDojo's standard observation gives the commanded one,
    so ORA's RoboDojo numbers are not on the official observation.
  • In free play the agent's bash runs as you, with credentials removed and no sandbox; code mode is the sandboxed
    way to script.
  • ORA installs as a Git checkout; there is no PyPI package yet. macOS runs the CPU LIBERO path only.

Website · Documentation · Blog · Gallery · Changelog

ORA preview (October 2026)

Pre-release

Choose a tag to compare

@freemty freemty released this 18 Nov 23:49

The first public tree of ORA, as published in October 2026 and described in the October post: open-weight models run LIBERO, RoboTwin and RoboDojo tasks through ORA's tools, from one command on one GPU.

It carried no tag when it was published; this tag keeps it reachable now that main has moved on to v0.1.0. To stay on it:

git clone https://github.com/fast-rua/ORA.git && cd ORA
git checkout v0.1.0-preview

and follow the README of that tree. It receives no further changes; the changelog says what an upgrade to v0.1.0 needs.