A robotics research workbench for simulating, observing, and evaluating multi-robot autonomy in MuJoCo.
Argus connects robot simulation, perception, mapping, exploration, and locomotion to a browser dashboard. Its goal is to make algorithms easier to compare under shared scenarios and measurable conditions.
| Area | Implemented |
|---|---|
| Simulation | Multiple Unitree Go2 robots in a shared scene, RGB/depth cameras, scene selection, and interactive spawn placement. |
| Mapping and exploration | ICP baseline, pluggable SLAM backends, voxel map merging, frontier exploration, A* planning, and region-biased coordination. |
| Perception | Configurable detectors, depth-based 3D boxes, object tracking, semantic maps, and detection metrics. Optional models require setup. |
| Browser dashboard | Live cameras, point clouds, voxels, meshes, trajectories, navigation controls, and a visual pipeline editor. |
| Locomotion evaluation | Analytical trot baseline, seeded terrain and disturbance scenarios, tracking/stability metrics, and structured exports. |
| Extensibility | Protocols and registries for backends, robot platforms, and controllers; MCP tools for simulation queries and commands. |
Scope: Simulation only, with CPU-oriented defaults. Map fusion uses simulator ground-truth poses as a shared reference; this does not establish localization robustness on hardware. Go2 assets are included. The AGIBOT X2 adapter requires external licensed assets and a walking controller. Learned locomotion, MPC, and WBC remain unimplemented controller placeholders.
Stack: Python, MuJoCo, Open3D, FastAPI, React, TypeScript, Three.js, and React Flow.
Requires uv and Node.js 22.12+.
uv sync --python 3.12 --extra web --extra dev cd frontend npm ci cd .. uv run argus --scene flat --static
Open localhost:8000. The frontend builds at startup. Remove --static to enable motion, or use --pick-spawn to choose a scene and starting positions in the browser. On PowerShell, use npm.cmd ci if script policy blocks npm.
See the usage guide for office assets, optional detectors, CLI options, and development commands.
uv run argus eval-locomotion --controller analytical_trot --scenario flat_ground --seed 101 --seed 202
Runs write manifest.json, steps.jsonl, episodes.csv, summary.json, and comparison.md under outputs/locomotion-evals/. The benchmark guide covers scenarios, metrics, and comparison matrices.
These are research directions, not shipped features:
- Experiment comparison: Browse saved runs, replay failures, and compare results across scenarios and seeds.
- Reusable robot data: Record synchronized observations and actions, then add dataset import/export adapters.
- Learned control: Evaluate residual and direct policies against the analytical baseline before adding training workflows.
- Robust coordination: Study sensor noise, pose uncertainty, communication limits, and larger teams.
- Persistent world models: Connect the existing belief-graph prototype to live observations and test its value for task planning.
Next proposed milestone: multi-robot exploration evaluation plan.