April 2026 demo rides — the results
In April we opened demo rides to the public across the whole of Tartu. Between 14 April and 28 May 2026, we completed 19 rides with passengers in the back seat. Passengers chose their own pickup and drop-off bus stops, so the routes were not ours to pick.
We recorded every ride, annotated everything the car did wrong, and asked the passengers what they thought.
The headline numbers
| Rides | 19 |
| Recordings (individual legs) | 88 |
| Distance driven | 323 km |
| Time driven | 13 h 57 min |
| Safety driver disengagements | 151 |
| Distance per disengagement | 2.14 km |
| Time per disengagement | 5.5 min |
| Autonomy, by distance | 97.8% |
| Autonomy, by time | 95.1% |
21 of the 88 legs, totaling 65 km, were driven without a single disengagement. The longest fully autonomous leg was 7.5 km and lasted 16 minutes. The rides were split between two safety drivers, whose km-per-disengagement rates differ by 5%, so the number reflects the software rather than the person supervising it.
How this compares to our earlier campaigns
| Tiksoja 2022 | Tiksoja 2023 | Tartu 2025 | Tartu 2026 | |
|---|---|---|---|---|
| Distance, km | 331.2 | 322.6 | 312.8 | 323.1 |
| Time, min | 704 | 601 | 899 | 837 |
| Disengagements | 263 | 199 | 217 | 151 |
| km per disengagement | 1.26 | 1.62 | 1.44 | 2.14 |
| min per disengagement | 2.7 | 3.0 | 4.1 | 5.5 |
| Autonomy by distance | 93.2% | 91.7% | 96.8% | 97.8% |
| Autonomy by time | 85.9% | 90.6% | 94.1% | 95.1% |
Each demo ride campaign covered roughly 320 km, but the software and driving environment differed a great deal across them. 2022 ran on Autoware.ai; since then, we have run increasingly capable versions of Autoware Mini. The first two were on the Tiksoja on-demand route, mostly rural roads with light traffic. 2025 covered Riia street and the city center. 2026 covered the whole of Tartu, so the last column is both the newest software and the most challenging driving of the four.
One caveat on the first two columns. In 2022 the safety driver stopped the car manually at 55 bus stops as a matter of policy rather than because anything failed, and both years include manually driven turnbacks. Corrected for those, 2022 and 2023 land almost on top of each other at about 1.75 km per disengagement, so the apparent jump between them is mostly due to a change in what was counted.
What we count as a problem
Disengagements are the standard metric in autonomous driving, but they don’t accurately reflect the end-user experience. A car that brakes hard for a bush at the roadside does not need the safety driver to take over; it is just unpleasant to ride in. So we annotated every noticeable misbehavior. That came to 662 events across the 19 rides:
| Event type | Count | Share |
|---|---|---|
| Unnecessary or excessive braking | 349 | 52.7% |
| Disengagement (safety driver took over) | 151 | 22.8% |
| Unnecessary swerving | 102 | 15.4% |
| Other | 36 | 5.4% |
| Missing or late braking | 19 | 2.9% |
| Manual start (car had to be driven out of a spot) | 5 | 0.8% |
Each event is traced to the module and the specific node in Autoware Mini that caused it.
Which part of the stack is at fault
Every one of the 662 events is attributed to a module. The right-hand column narrows the same data to the 151 cases where the safety driver took over.
| Module | Share of all 662 events | Share of the 151 disengagements |
|---|---|---|
| Local planner | 282 (42.6%) | 55 (36.4%) |
| Perception | 249 (37.6%) | 42 (27.8%) |
| Map | 92 (13.9%) | 20 (13.2%) |
| Human operator mistake | 9 (1.4%) | 9 (6.0%) |
| Remote assistance | 9 (1.4%) | 8 (5.3%) |
| Controller | 10 (1.5%) | 7 (4.6%) |
| Localization | 6 (0.9%) | 5 (3.3%) |
| Other | 5 (0.8%) | 5 (3.3%) |
Planning and perception account for four out of five events. Map problems come third and are the cheapest to fix: move a lane centerline, add a stop line, and the problem is gone for every future ride.
The individual problems that cost us autonomy
These are the eight most common causes, accounting for 56 of the 151 disengagements. The remaining 95 are spread across many smaller categories, none with more than four.
| Problem | Disengagements |
|---|---|
| Approaching vehicles detected too late | 12 |
| Double traffic light classified incorrectly | 11 |
| Map-based prediction too short | 7 |
| Pedestrian detected too late at crosswalk | 6 |
| Pedestrian standing at a crosswalk not handled | 5 |
| Our car blocked by an obstacle in its lane | 5 |
| Predicted collision blocks our car at an intersection | 5 |
| Localization problems | 5 |
Approaching vehicles detected too late almost always happened during unprotected left turns. The car looks at the oncoming lane, sees nothing, and decides it can go — but another vehicle is approaching fast, just beyond the range at which our lidar clustering can resolve it. The safety driver stops the maneuver before it starts. This is our largest source of disengagements and the most safety-relevant one.
Double traffic lights are pairs of lights mounted side by side, one for the turning lane and one for the lane going straight. Our YOLO detector fails to reliably distinguish them and misreads the neighboring lane’s red as ours, so the car stops at a green light and stays there. Eleven of the 22 cases ended in a disengagement because we didn’t want to block traffic behind us.
Pedestrians at crosswalks failed in both directions. Six times the car detected a person crossing too late; five times someone waiting at the curb, clearly about to cross, was not treated as a reason to stop. In the other direction, it was too timid: 12 times, with our light green and the pedestrian’s red, someone merely walking up to the crosswalk made the car brake. Another 29 times a stationary roadside object was perceived as moving, had its predicted path extended across a crosswalk, and slowed us for no reason.
Braking and swerving for nothing
Most events do not require an intervention from the safety driver, but you still feel them from the back seat. These are the eight most common problem categories across all 662 events, together accounting for 246 of them:
| Problem | Events |
|---|---|
| Roadside object cluster expands onto the path | 42 |
| Path extends into the opposite lane while swerving | 38 |
| Lane centerline mapped too close to the boundary | 34 |
| Stationary roadside object activates a crosswalk | 29 |
| Lane boundary position inaccurate | 29 |
| Inaccurate radar detection from an oncoming vehicle | 27 |
| Tracker assigns an object a wrong speed and moves it onto the path | 25 |
| Double traffic light classified incorrectly | 22 |
Most of this braking is caused by objects that are not really in our way, but are briefly believed to be. A parked van’s lidar cluster extends a few centimeters into the safety buffer, or a radar return from an oncoming car appears in our own lane behind a curve.
The swerve planner is the node most often responsible for a misbehavior, with 120 logged events. Its most common failure is subtle: where the lanes are narrow, the safety buffer around the local path reaches slightly into the oncoming lane. The buffer then intersects an approaching vehicle that is entirely within its own lane, and the car brakes hard for traffic that was never going to hit it.
What the passengers said
Thirteen riders filled in a questionnaire after their ride, eleven in Estonian and two in English, rating five aspects of the experience from 1 to 5.
Thirteen self-selected responses do not make up a proper survey, but the pattern is clear. People felt safe and still rated the car well below a human driver. Eleven of the thirteen chose 4 or 5 on safety and every respondent said the ride met or exceeded their expectations. On the last question, where 3 means comparable to a competent human on the same route, ten put the car below that and only two above it.
What separates the machine from the human is smoothness. The riders who named a specific rough moment identified some of the most common braking events:
“An unexpected cloud of dust caused sudden braking.”
“Some oncoming cars caused unexpected slowing, but overall everything was very OK.”
A dust cloud read as an obstacle is a lidar cluster that isn’t really there, and oncoming cars causing unexplained deceleration is the radar giving inaccurate detections.
One rider raised something that we had not counted:
“Because the car drove 40 km/h and stopped at every uncontrolled intersection, I was worried that some other car would drive into us.”
We treat excessive caution as a comfort cost; this rider treats it as a safety risk. Another passenger gave a low rating due to a missing capability rather than an error: the car does not change lanes, which forces longer and stranger routes.
Asked what made the ride feel the way it did, riders kept naming the screen showing what the car perceives and the crew explaining what was happening.
“It felt like a private tour arranged especially for me. Looking at the tablet feels like taking part in a simulation, but the route is my everyday commute and we were really outside.”
That qualifies every number above. What we measured is a guided demonstration with two engineers in the front seats answering questions, not an unattended robotaxi ride. The overall feeling score averaged 6.3 out of 7, and some of that is the crew rather than the software.
(Quotes translated from Estonian.)
Human factors
Nine of the 151 disengagements were our own mistakes rather than the car’s: engaging without a global path (4), engaging in the wrong gear, disengaging before the recording was stopped, one accidental disengagement, a phone that fell onto the emergency stop button and cut power to the drive-by-wire system, and a passenger who left a door open.
Four more came from passengers changing their minds mid-ride, including two elderly ladies whose actual destination was inside a parking lot, which we have not mapped and cannot drive into.
Remote assistance
Remote assistance was in use for almost the whole campaign: an operator connected over the network could confirm that the car could enter an intersection or draw a path around an obstruction. Eight disengagements occurred when the vehicle received outside assistance. These can be divided into three categories: a drawn path with too sharp a turn (4), giving permission to enter an intersection despite oncoming traffic (3), and failing to check for an overtaking vehicle (1).
Getting stuck behind an obstacle was a top issue in earlier years and has dropped sharply in this year’s disengagement list — the operator can draw a path around an obstruction before the safety driver has to take over.
What we are fixing next
- Swerve planner — 120 events, the largest single source in the log. Stop the local path from reaching into the opposite lane, and stop unstable clusters from producing jerky swerving.
- Lidar clustering — 98 events, and the origin of our biggest disengagement cause. Detect approaching vehicles earlier, and stop roadside clutter from expanding onto the path.
- Pedestrian crosswalk checker — 77 events. Handle people who are standing and clearly waiting; stop reacting to people and objects that are not crossing.
- Traffic light detection — the double traffic light configuration, 22 events and 11 disengagements, all from one narrow failure mode.
- Lane changing, the missing capability our passengers noticed fastest, and the reason some routes are longer than they need to be.
- Map corrections — 92 events and the cheapest fixes available: centerlines too close to boundaries (34), inaccurate boundary positions (29), bus stops missing from the map, unmapped road closures and speed bumps.
Come ride with us
Demo rides are still open. If you would like to contribute a route we have not driven yet, preferably a difficult one, you can book a demo ride.