bagio — ROS1 bag access without ROS

A small Python 3.11 library for reading recorded ROS1 .bag files without any ROS installation, built on the pure-Python rosbags library. It is the shared bag-access layer of the AI4MAP offline pipelines (offline_mosaicing, offline_sonar_processing).

Install (editable, from a checkout):

pip install -e .

Modules

• typestore.py — message-type registration. Types are registered from each bag's embedded message definitions (connection.msgdef.data), not from rosbags' bundled stores: the bundled ROS1 store lacks some standard types (e.g. tf2_msgs/TFMessage) and knows nothing about cola2_msgs, and the embedded definitions stay correct even if a message definition changes between campaigns. stamp_to_ns() converts header stamps to integer nanoseconds.

• bag_reader.py — BagSet opens one or more bag parts (*_0.bag, *_1.bag, …) with a shared typestore and yields deserialized messages merged in timestamp order. decode_image() converts sensor_msgs/Image (bgr8/rgb8/mono8/bayer, row padding handled) and CompressedImage payloads to BGR numpy arrays.

• nav_series.py — the cola2_msgs/NavSts stream as sorted arrays with nearest-sample and interpolated access. Yaw interpolates on an unwrapped copy and re-wraps, so the ±π seam cannot produce a bogus midpoint. median_filtered(field, window_s) returns a centered sliding-median copy of a field (e.g. DVL altitude bottom-lock hop removal; refuses yaw), and interp_values(t_ns, values) linearly interpolates a precomputed per-sample array (e.g. the output of median_filtered) at an arbitrary timestamp. A mission has one physical AUV trajectory; every sensor that rode it consumes one GPS-free correction solved from the camera pose-graph. load_trajectory(path) loads and validates a trajectory_refined.json correction artifact — checking schema_version, status, a "relative" frame, and a keyframe list matching n_keyframes — and raises ValueError if any check fails. apply_trajectory_correction(nav, artifact, origin_tol_m=1.0) returns a corrected copy of a NavSeries: the artifact's per-keyframe delta_n/delta_e/delta_yaw are interpolated at each sample's timestamp (linear between keyframes, held flat beyond the first/last) and added to north/east/yaw (yaw re-wrapped to (-π, π]); the artifact's NED origin must be within origin_tol_m metres of the nav origin, or it raises (guards against a mis-paired artifact). NavSeries.with_correction(path, origin_tol_m=1.0) is a convenience wrapper for the two calls.

• tf_buffer.py — transform tree from /tf_static + /tf. lookup(target, source, t_ns) composes the chain at an explicit timestamp, interpolating dynamic edges (linear translation + slerp rotation). Queries beyond max_extrapolation_ns (default 0.5 s) outside an edge's recorded range raise, mirroring tf's ExtrapolationException. A frame reparented mid-recording raises rather than silently discarding history.

Numerical note

All time interpolation runs on nanosecond offsets from the first sample of each series, never on absolute epoch nanoseconds: at 2026-epoch magnitudes (~1.8e18 ns) float64 resolves only ~256 ns, which can collapse distinct timestamps (and, e.g., make scipy Slerp reject tied knots). Offsets are exact in float64 for any realistic mission length.