Kai Team
Natural language driven robotics simulations in Mujoco
Core trick
A model writes MuJoCo scene specs from an asset catalog and dispatches collision-checked Franka arm motions through a motion controller.
Description
Natural language driven robotics simulations in Mujoco
README excerpt
# Astra Robot World Astra composes MuJoCo scenes from an extensible asset catalog and directs a Franka Panda arm to stack, pack, and sort small objects, with collision-checked motion. Physics and the native viewer run locally on macOS; live language requests use your configured API. Explore the [interactive architecture canvas](web/architecture.html) or [architecture reference](docs/architecture.md) for Astra’s role, all 25 tools, the motion controller, experiments, and storage. In the running app, click **Architecture**. ## Run on your Mac Prerequisites: Git and [uv](https://docs.astral.sh/uv/getting-started/installation/). ```sh uv sync --locked uv run python scripts/fetch_assets.py ./run.sh ``` Open **http://127.0.0.1:8765**. The browser embeds actual MuJoCo rendering alongside chat, Action Lab, Stop, and manual physics controls. The default launcher uses one simulation process and a fixed port; invoking it again reuses the existing app. **Include current simulation view** is enabled in chat by default. Each message sends the displayed simulation image alongside your text, labeled as live, experiment, or action replay. Disable it for text-only requests. Images are limited to 960 × 720 pixels in the browser and are removed from conversation history after that turn; they are not written to disk. If the renderer is unavailable, chat sends text only. Vision adds API image usage and requires an image-capable configured model. The attachment is a snapshot, not a continuou