Data infrastructure for robots.

A pandas-like analysis workbench for ROS 2 (MCAP) bag files — the data platform for what robots already recorded. No ROS installation required.

• Bounded-memory streaming core: chunked CDR decode, running aggregators, and tolerance-bounded sync buffers analyze multi-GB logs with memory bounded by chunk size, not file size — enforced by a dedicated memory-regression CI suite.
• Bags → training data: streaming exporters to 7 analytics/ML formats (Parquet, HDF5, Zarr, LeRobot, RLDS…) with multi-stream time synchronization verified by eager/streaming equivalence tests.
• Data quality as a first-class layer: streaming health checks, declarative bag contracts usable as CI gates, and self-contained incident reports.
• Notebook UI: every capability is a cell — plot, brush a window for a grounded anomaly explanation, free-form Polars queries, cross-run comparison, local CLIP semantic frame search, one-click export.

Replay-first observability for robot fleets: watch a live fleet stream policy decisions, replay any decision through real physics, and see what the policy believed — and what it would have earned.

• Deterministic counterfactual replay: any logged decision re-rolls through closed-loop MuJoCo physics, so "what if" is answered with dynamics, not vibes.
• Evidence-tiered fragility search and physics-gated policy promotion — an improvement flywheel whose gate blocked its own first candidate, reported straight.
• Cognition layer: a Q-ensemble with a calibration-gated kill switch, plus deterministic incident narratives with portable evidence packets.
• Two morphologies, one contract: a Unitree G1 humanoid and a Go1 quadruped run through the same robot-adapter interface — connect your own via the capture daemon.