Skip to content

Why did the robot do that?

A robot stops on your line. Finding out why pulls your best engineer off the roadmap — hours of triage, often a trip to site, while the line stays down. Torsen reconstructs the incident, pinpoints where it first went wrong, and hands support a case they can act on.

Start a pilotor just keep me posted

Every robot incident burns money in four places.

Not in missing data — you already log everything. In the human hours it takes to explain what happened.

$36k–$2.3M

per hour of downtime

FMCG to automotive · Siemens 2024

8h+

senior-engineer time per incident

pulled off the roadmap

$1.5k–$6k

per emergency site visit

when remote triage can’t answer it

$6.9M

service-ops loss, one RaaS vendor

in a single year · Symbotic FY2025

A 300-robot fleet leaks roughly $400k a year to this. Torsen is built to take back a third of it.

Same failure. Two ways to find out why.

A learned-policy humanoid lost its footing in a shared workcell and toppled toward an operator; the e-stop fired. Nobody was hurt. Scrub the raw log to find out why — then reconstruct it with Torsen.

Why did it fall toward the operator?

Raw rosbag timelineWhat you’d scrub today
/tf
/tf_static
/joint_states
/imu/data
/odom
/cmd_vel
/policy/action
/policy/value
/footstep_plan
/com_state
/foot_contacts
/camera/color/image_raw
/camera/depth/image_rect
/perception/tracks
/scan
/diagnostics
/estop
/clock
A real recording, replayed flat — dozens of topics, no culprit marked.
Torsen reconstructionThe few signals that mattered

Freezes your clock, jumps to the failure window, surfaces the few signals that mattered.

Representative reconstruction, rendered — synthetic data, not a real customer incident. Built to show what a Torsen reconstruction looks like. The left clock is your own time, scrubbing right now. The right is a representative Torsen reconstruction time.

One reconstructed case. Not a log dump.

Torsen finds the failure window in a recording you already have, lines the run up against the runs that worked, and marks the first moment it left them. Support gets a case they can read, act on, and hand to anyone.

Torsen · incident replayrepresentative · synthetic data
The Torsen incident replay: an AMR halted at a pick station, four synchronized signal lanes each shaded with the envelope of fourteen known-good runs, the trace turning amber where it leaves that envelope at the marked first divergence, and a diagnosis panel naming the likely cause.
Torsen · case index
The Torsen fleet case index: incidents-today stat, per-robot sparklines, and a table of open incidents across the fleet.

Compared against known-good

The same task, run normally, is the baseline — the band each signal should have stayed inside.

First divergence

The earliest moment the run left that band — found before the stop anyone would notice.

A shareable case

Diagnosis, evidence, and next action — one record support can act on and hand to anyone.

Keep your recorder. Torsen begins where capture ends.

Your stack already holds the recording, freezes the buffer when the fault fires, and assembles ROS, PLC and safety state into one view. None of that tells you which signals mattered, or where the run first left nominal. That part is still done by hand — by your most expensive engineer.

Your stack captures

  • rosbag2 · MCAP

    The recording, already on disk.

  • Freeze-frame · rolling buffer

    The fault doesn’t start the recording — it stops it.

  • SOVD · ros2_medkit

    Faults, snapshots and PLC state, assembled and served over HTTP.

  • Foxglove · Rerun

    Replay and visualization — once you know where to look.

Torsen investigates

  • Finds the failure window

    In the recording you already have — no new instrumentation.

  • Compares against known-good

    The same task, executed normally, is the baseline.

  • Marks the first divergence

    And separates the cause from the cascade that followed.

  • Hands back a case

    Evidence, next action — and what was never recorded.

Assembly is architecture. Interpretation is the work.

Read-only, start to finish — Torsen never writes to your bag, your robot, or your ticket.

You pay because it closes more cases with fewer experts.

Triage

Todayhours before a first useful hypothesis

Torsenthe case is already built when you open it

Expert escalations

Todayyour best engineer, pulled into every hard case

Torsensupport closes it from the record

Site trips

Today$1.5k–$6k when remote triage can’t answer

Torsenthe reconstruction answers remotely

Recurrence

Todaythe same failure comes back unnoticed

Torsena fix bound to a version, watched for regression

OEM example

144 incidents a year leak about $399k in senior hours, support, and trips. Cut that by 30% and Torsen returns roughly $120k a year — before churn or SLA credits.

Bring us your hardest unexplained failure.

We take a handful of incidents at a time. Tell us about a failure your logs couldn’t explain — one you already have a recording of — and we’ll follow up to reconstruct the grounded why with you, founder-to-engineer.

One failure your recordings show, but never explain. No customer names needed.

Do you have a recording of it? (required)

A rosbag/MCAP of the failure window is ideal, but not required to apply.

Three questions, then a founder reads it and replies in person — we’ll ask about your stack there, not here.

Your answers go to one inbox via Formspree — no marketing, never shared.

Not ready to bring a failure yet?

One email when there’s news. No spam.