ArcheOne-AI/RoboScript

What it is

RoboScript is an agent framework for robot manipulation that does not train an action policy. A frozen foundation model starts from one demonstration, rehearses the task in a world such as a simulator, and stores what it learns in a Task Memory in context rather than in weights. The project calls this World In-Context Learning (WICL), described in its arXiv paper.

The framework has four stages: build reusable skills from demonstrations, learn a Task Memory from one demonstration, deploy the frozen memory on new scenes using only real-robot-observable sensing, and improve through reflection on past evaluations. It is currently validated on dual-arm tabletop manipulation.

Results reported

On RoboTwin 2.0 (50 bimanual tasks), the authors report 83.4% mean success with one demonstration per task and no gradient updates, compared with 61.7% for the strongest trained policy using 50 demonstrations per task. Rehearsal happens only on clean scenes; the authors state randomized scenes show almost no drop (83.6% to 83.2%). Additional benchmark results are described as in progress.

Getting started

The built-in Mock benchmark runs one episode without a simulator, GPU or model. Requirements: Python 3.10+ (per the badge) and uv.

git clone https://github.com/ArcheOne-AI/RoboScript.git
cd RoboScript
uv sync --all-packages
cd examples/tutorials
uv run python -m quickstart_mock.run

After the mock run, continue with the tutorials, pick another benchmark, or read the documentation. The vocabulary defines terms such as skillset, brain and recipe.

Included capabilities

The repository contains the framework core, three learning methods (wicl, rsi, skill_builder), the vision service, and a real-robot package. Benchmarks:

Benchmark	Status
Mock	Built in, no simulator needed; used by tutorials
RoboTwin 2.0	Full pipeline with official evaluation
BEHAVIOR	Integration in development

Methods, benchmarks, embodiments and skillsets are plugins. You can add a sim benchmark, skillset, real robot or other extension without changing the core.

Infrastructure

Each component runs as its own service. The model gateway routes LLM calls to any provider with session pinning and failover; the agent plane is limited to what a real robot can sense; the control plane runs the learning protocol and holds ground truth, with only the official check deciding success; services cover motion planning (cuRobo), perception (GroundingDINO, SAM 3) and blob storage.

Constraints

The project is pre-1.0, so interfaces may change. Real-robot deployment code is still listed as a TODO, although the README reports memories rehearsed in simulated ALOHA running on a dual-arm PiPER through parameter remapping. Benchmark simulators and their assets are not part of the repository and carry their own licenses. The code is Apache-2.0, with third-party adapted code listed in NOTICE.

HotRepo's take

RoboScript was listed on the HotRepo new-project board for the Python language on 2026-11-12, with 7 stars gained in a window of about 1.5 hours and 110 stars total at the end of that window. That is a single early record rather than a trend: the project appeared on GitHub only in November 2026 and has no all-language daily rank recorded yet. For readers, the more meaningful signal is the reported one-demonstration, no-gradient setup and the mock tutorial path, which makes the framework testable before any simulator or GPU is involved.