The robotaxi went backward, and so should your roadmap math

Physical AI is shipping. Robots carry totes in warehouses. Robotaxis pick up paying riders in six cities. The demos are real. But the gap between what these companies promised and what they delivered just got measurable, and the numbers are not moving the way the pitch decks said they would.

If you run product or design and you set your roadmap against big AI promises, this quarter is a useful mirror. Let me catch you up on what actually changed.

The line on the chart went the wrong way

Tesla put out a chart that looked like steady growth. Then people did the math. Because the numbers were cumulative, they hid the real story: Tesla's paying robotaxis covered about 1.1 million miles in Q1 and only 700,000 in Q2, a 36% drop. This was after expanding to six cities. More cities, fewer miles.

Compare that to the field. Tesla's total for the year is less than 10 percent of Waymo's weekly mileage, and Waymo drives around four million miles a week. One BNP Paribas analyst called Tesla's trend "an alarming decline." The stock fell more than 13% the next morning.

The lesson is plain. When a vendor shows you a curve that only goes up, ask what the axis actually measures. Cumulative charts hide flat quarters.

New cities are a headline, not a fleet

Tesla announced Orlando and Tampa the day before earnings. It did the same with Dallas and Houston before the last earnings call, then never scaled them. As of late July, the Robotaxi Tracker showed 16 unsupervised vehicles in Austin, four in Dallas, and one in Houston. Numbers go up and down with no explanation.

Adding a city looks like progress to investors. It is cheap to announce and hard to verify. Meanwhile wait times are long and vehicles pick up in the wrong spots, because supply is thin.

Watch for this pattern on your own roadmap reviews. A new logo, a new market, a new integration on the slide does not mean the thing is deployed at scale. Ask how many units are actually live, not how many places got an announcement.

The story about why it's slow keeps changing

Here is the part worth flagging for your next review. Tesla spent years saying its 10 million customer cars were silently collecting data to train future robotaxis. This quarter Musk said the company needs to accumulate driving data specific to the Cybercab before it can put many on the road. That is a different story than the one told for years.

The reason for the delay also shifted. For years the block was "regulation," though the company never named which rules. Now the block is "proving safety." When the explanation for a miss keeps moving, the miss is the real signal.

Apply that to your vendors and your own team. If the reason a feature slipped changes every quarter, stop debating the reason and start pricing the delay.

Smoke, sirens, and the edge cases that aren't edges

Safety is where physical AI stops being a slide. Zoox shipped a software recall to all 105 of its vehicles after one robotaxi got confused by heavy smoke at a fire scene and braked hard before a teleoperator reversed it out. A regulator had just warned AV companies to stop interfering with first responders, calling the failure to handle emergency scenes "a functional insufficiency."

The key phrase in that warning: emergency scenes "are not rare or extreme edge cases." Waymo had at least six incidents where responders had to physically move a car. Power outages have stalled Waymo fleets more than once, including a one-hour San Francisco pause and a July 4 jam that pushed the mayor to demand tougher rules.

The messy real world is not the exception you patch later. It is the product. Budget for it up front, not in a recall.

The deep cut

One company is playing this differently, and it's the tell. Agility Robotics has a humanoid, Digit, that already earns revenue moving totes for Amazon and Toyota, with $300 million in contract orders. Its co-founder said the safety stack does not run through generative AI: "You don't want to get creative with your safety stack." They use AI for scale, humans-only zones for safety, and they refuse to promise in-home robots yet.

That is the split that matters for your roadmap. Agility narrowed the promise to what it can ship safely and let the demos stay boring. Tesla kept the promise huge and let the numbers slip. The teams that win the next two years will be the ones who separate the part that can be creative from the part that cannot fail, and who resist the pull to announce the big version before the small one works.

Three questions for your team

• When our roadmap shows growth, is the chart cumulative or per-period? If a competitor's chart went backward the way Tesla's did, would ours catch it?
• Which of our "edge cases" are actually core cases, like smoke or an outage, and are we funding them now or planning to patch them after an incident?
• Where are we announcing scope we haven't shipped? For each new market or integration on the deck, how many units are actually live?