Work In Progress


A fully autonomous vehicle functions without human intervention. OAS proposes measuring autonomy progress in terms of Miles Per Intervention (MPI). MPI not only captures the industry-accepted metric of Miles Per Disengagement (MPD), but also tracks any time a human is required to make a safety critical decision (e.g. remote tele-operations).

Field Collection

Classifying Interventions

Interventions are grouped into the following categories:

CriticalDisengagement where driver intervention is necessary to prevent a safety-critical or reckless situation. MPCI
System FaultDisengagement where the vehicle experiences a hardware or software failure and responds appropriately. MPFI
MalfunctionDisengagement where the vehicle behavior is not working as designed. MPMI
UnsupportedDisengagement where the vehicle encounters a scenario that is not yet supported. MPUI
ExperienceDisengagement due to uncomfortable riding experience. MPEI
RemoteIntervention from remote vehicle operator. MPRI


We track the average number of miles by different classifications of interventions.

Miles Per Intervention (MPI)

MPI = (Total Autonomous Miles) / (Total Interventions)

Miles Per Critical Intervention (MPCI)

MPCI = (Total Autonomous Miles) / (Total Critical Interventions)

Additional Notes: Since drivers disengage early for safety-critical scenarios, any disengagement flagged as such should be re-simulated to determine the outcome of the situation.

Miles Per System Fault Intervention (MPFI)

MPFI = (Total Autonomous Miles) / (Total System Fault Interventions)

Miles Per Malfunction Intervention (MPMI)

MPMI = (Total Autonomous Miles) / (Total Malfunction Interventions)

Miles Per Unsupported Intervention (MPUI)

MPUI = (Total Autonomous Miles) / (Total Unsupported Interventions)

Miles Per Experience Intervention (MPEI)

MPEI = (Total Autonomous Miles) / (Total Experience Interventions)

Miles Per Remote Intervention (MPRI)

MPRI = (Total Autonomous Miles) / (Total Remote Interventions)

Collection Cycle

Field Collection

Data collection begins in the field with the drivers of the Operations team. The primary method used to identify software problems is tracking disengagements. Every disengagement is geotagged and time-stamped. Operations Specialists can provide additional context for the disengagement.



Every trip is recorded automatically and our riders are encouraged to leave feedback. Our data is stored locally on the vehicle and syncs at night to our servers where it is processed, evaluated, and archived. With the uploaded data, our engineers can select the trips to replay and visualize the content at any time.

Collected Data


Simulation plays a critical role in helping validate new software releases. It allows vehicles to accumulate miles driving challenging scenarios in virtual communities.

Whenever a disengagement occurs, it can be re-simulated on future releases of the autonomy software. Re-simulation allows the team to playback previously collected data and simulate how the vehicle would have reacted if the driver had not intervened.