Docs / Platform
Roadmap and scope
What Invariant builds, what it integrates and what it deliberately doesn't build, plus the order it ships in.
Design preview. These docs describe how Invariant is being built. Commands, output and limits are illustrative until release, and stage tags mark what ships first. See the roadmap.
Build, integrate, don't build
| Invariant builds | Invariant integrates | Invariant doesn't build |
|---|---|---|
| The test spec and runner | Simulators (Isaac Sim, Gazebo, MuJoCo, O3DE, custom) | A simulator |
| Scheduling, caching, sharding, budgets | Compute (your machines, cloud GPUs) | A GPU cloud |
| Paired statistics and verdicts | Log formats (MCAP, rosbag2) | A proprietary log format |
| Run history and run records | Viewers for inspecting runs | A 3D viewer |
| Metric packs and custom metric SDK | CI systems, chat, coding agents | A CI system |
| Scenario variation and coverage | World and asset tools | Photoreal world generation |
| Compatibility matrix, tested nightly | ROS 2 distributions, middleware | A robotics framework |
MVP: first pilot release MVP
Focus: log replay for localization and perception, on the customer's machines. No simulator choice needed, no GPU fleet required.
invariant.yaml(short and full form),init,doctor,run,scenario add,compare,export- Local Docker runner and self-hosted runner agent
- ROS 2 adapter (Humble, Jazzy, Lyrical) and process adapter
- MCAP and rosbag2 replay with slicing, warm-up and ground truth
- Localization and perception metric packs, custom Python metrics
- Paired runs, confidence intervals, tolerances and four verdicts
- CI integration through exit codes and Markdown summaries
- Minimal web report: suite results, history per scenario, links to MCAP files
Next: after the first pilots Next
- Simulation scenarios: Gazebo, Isaac Sim and MuJoCo adapters, scenario variation, coverage
- Navigation and manipulation metric packs
- Invariant cloud workers (CPU, L4, L40S) with image caching and warm simulators
- Sequential testing, false discovery rate control, flaky detection
- Pull request app, chat alerts, bisect, MCP server,
planandreplay - Python adapter for in-process model evaluation
Later Later
- Hardware rig runners and rig queues
- Workers in the customer's cloud account; fully self-hosted deployment
- Release evidence reports
- SSO, SCIM, SOC 2
- Open-source release of the spec and local runner
- Measured agreement between sim and real results per scenario
This roadmap reflects current plans and pilot feedback. Order and scope will change as pilots teach us what matters most.