estimator_offline
Overview
Two workstation tools for looking at the state estimator without a car. estimator_sim records a simulated drive: what mti610_bridge and bd992_bridge would have published, on the same keys and schemas with arrival jitter, plus the truth. estimator_offline runs the estimator over any bag, simulated or real, twice over: once as the live node would have seen it, and once as a whole-drive solve. Both write ordinary bags, so scope opens the result.
estimator_sim ─► sim bag ─► estimator_offline ─► estimate bag ─► scope (configs/scope/estimator.yaml)
▲
bag record (a real drive) ──────┘
Neither ships. They link the simulator and hold a whole drive in memory, which is why they are under tools/.
Running it
./build/tools/estimator_sim/estimator_sim --out /tmp/skidpad --scenario skidpad --outage 25:30
./build/tools/estimator_offline/estimator_offline --in /tmp/skidpad --out /tmp/skidpad_est --time-offset 0.028
./build/apps/scope/scope --config configs/scope/estimator.yaml --bag /tmp/skidpad_est
estimator_sim prints the --time-offset its recording needs as its last line: the simulated GNSS latency floor minus the IMU’s. For a real drive, use the value calibrated in configs/state_estimator/state_estimator.yaml.
estimator_sim
| Option | Default | |
|---|---|---|
-o, --out <dir> | required | Bag directory to write. |
--scenario <name> | skidpad | skidpad, spin, figure8, parked, track or stopandgo. |
--drive <s> | 40 | Seconds of driving after the start. |
--seed <n> | 1 | Sensor-noise seed. |
--outage <from:to> | none | A GNSS outage, in seconds from the start. |
The bag holds nodes/mti610/mtdata2/delta_q and delta_v at 100 Hz, the GSOF records the estimator reads under nodes/bd992/gsof/ at 10 Hz (GSOF 12 and 38 at 1 Hz, as the shipped bd992.yaml sends them), and sim/truth as a VehicleState at 100 Hz, so it plots against the estimate field by field.
Every scenario starts parked for 5 s, because the estimator levels itself on a stationary accelerometer. skidpad then launches round a 40 m circle to 18 m/s and swings the slip angle between about 3° and 26° every 8 s; spin is the same and ends by rotating the nose past 90° of slip; figure8 drifts a figure of eight with slip up to about 29° that changes sign at the crossing; parked does not move. track laps a stadium at 22 m/s, straights run true and corners drifted at up to 20° of slip, which is what a mounting yaw is learned on; stopandgo drives, stops on a small grade for 8 s, and drives on at a new heading, which is what mounting roll and pitch are learned on.
estimator_offline
| Option | Default | |
|---|---|---|
-i, --in <dir> | required | Input bag directory. |
-o, --out <dir> | required | Output bag directory. |
-c, --config <file> | configs/state_estimator/state_estimator.yaml | The node’s config. |
--time-offset <s> | from the config | Overrides imu.time_offset_s. |
The output is every input message, unchanged, plus two tracks:
| Key | Schema | What it is |
|---|---|---|
estimator/fls | VehicleState | The fixed-lag estimate at IMU rate: what the car would have seen, through the same pipeline code the node runs. |
estimator/batch | VehicleState | The whole drive solved at once from the fixed-lag run’s keyframes, one state per keyframe (10 Hz). |
Exit codes: 0 when the batch solve converged, 1 when it did not, when the estimator never started (no dual-antenna heading, or no IMU), or on a read or write failure, and 2 for bad usage.
estimator_offline logs a summary: keyframes, states, resets and malformed messages for the fixed-lag pass; keyframes, iterations, initial and final cost, stop reason and refused factors for the batch.
Reading the result
configs/scope/estimator.yaml lays out three panels (sideslip, speed and yaw), each with sim/truth, estimator/fls and estimator/batch. On a real drive there is no sim/truth and those traces stay empty.
The batch track is the nonlinear equivalent of an RTS smoother: every state sees every measurement, before and after it. On a clean simulated drive it is several times closer to the truth than the fixed-lag track (see the numbers in the design note).
Across a GNSS outage the batch track draws a straight line. A keyframe is made per GNSS epoch, so the outage has none, and scope joins the two either side. The fixed-lag track does not stop: between keyframes the IMU carries the state forward at 100 Hz.
Batch states carry GPS time and the bag’s clock is the host’s. The fixed-lag pass has both for every state it produced, so the tool takes the median of host minus GPS time over those (one sample in fifty) and places each batch state at its GPS time plus that offset. On a recording with a large clock step partway through, the batch track will be shifted on one side of it.
Checking the output
Use scope_sample_stats rather than a screenshot to compare the tracks: it reports sample counts, drops and the minimum and maximum each trace actually received. On the 45 s skidpad above, truth had 4499 samples with a minimum sideslip of −25.78°, the fixed-lag track 4225 at −25.757°, and batch 373 at −25.758°, with no drops.