The Last Fifty Feet

Executive Summary

Autonomous vehicle fleets have accumulated billions of miles of operational data on how to drive. They have not solved how to communicate. A robotaxi is dispatched to a fixed coordinate and has no channel through which the person it was sent to collect can correct it. An analysis of internal Waymo rider feedback found that seven percent of riders were dropped anywhere from fifty feet to half a mile from their actual destination.¹ Peer-reviewed study of real-world robotaxi riders identifies the pickup and dropoff phases specifically as a recurring source of difficulty, alongside limited flexibility and prolonged waits.² Riders report vehicles driving past them entirely, or pulling over and departing again after judging the spot unsuitable.³

The driving stack is not implicated in these failures. Modern autonomy navigates to a coordinate with high precision. The failure occurs upstream, in the layer that decides which coordinate the vehicle is given. That layer is a geocoder today: it converts an address into a fixed point and passes it down, with no perception, no dialogue, and no capacity to revise.

Curbside replaces that layer with an onboard agent that interprets rider language, grounds it against the vehicle’s existing perception feed and map, and proposes a revised destination goal. The agent never touches steering, braking, or trajectory planning. It proposes, and the certified driving stack retains full authority to accept, defer, or reject. Because every autonomy architecture, modular or end-to-end, already exposes a goal interface, a single integration generalizes across fleets. The result is a reduction in failed rides and remote operator escalations, delivered as software on compute the vehicle already carries.

1. Market Structure

Commercial robotaxi deployment has passed the pilot stage. Waymo completed more than four million fully driverless rides in 2024 alone, bringing its lifetime total past five million,⁴ and now operates a US fleet in the low thousands of vehicles serving roughly half a million paid rides per week.⁵ May Mobility launched fully driverless commercial service in Peachtree Corners, Georgia in February 2025, having completed over 400,000 autonomy-enabled rides across eighteen deployments in the United States and Japan.⁶ Avride and Uber began commercial robotaxi service in Dallas in December 2025.⁷ By early 2026, services were operating across the United States, China, and the United Arab Emirates, with Baidu’s Apollo Go accumulating seventeen million rides.²

Analysts project the robotaxi market to reach $45.7 billion by 2030, growing at a compound annual rate of 91.8 percent.⁸ Goldman Sachs forecasts the global commercial fleet expanding from roughly 7,000 vehicles to approximately six million by 2035, a market worth more than $400 billion, with the United States alone accounting for $48 billion.⁹

The structure of this market matters as much as its size. It is fragmenting across many operators running incompatible autonomy stacks rather than consolidating behind one. Each operator faces the same rider-coordination problem, and each would otherwise solve it alone, inside a stack that serves only its own fleet.

2. The Coordination Failure

A coordinate specifies a point. An address specifies a building. Neither specifies where a person is standing. A rider assigned to 400 Market Street may be at the front entrance, a side door, a parking structure, or a loading dock the map never indexed. The vehicle arrives where it was sent. The rider is elsewhere. Neither party has a mechanism to resolve the discrepancy in real time.

The industry has documented this at length, including in its own patent filings. A USPTO filing on peer discovery in transactional mobile applications observes that imprecise pickup location estimates in autonomous ride-hail can require a series of interventions to resolve, and that the problem is particularly acute when a self-driving vehicle performs the pickup.¹⁰ A separate filing on rider drop-off experience acknowledges that autonomous systems weighing safety, traffic law, and other constraints will select pull-over locations that riders find counterintuitive, including stopping farther down the road or on the opposite side of the street.¹¹

Field reporting confirms the pattern across operators. Coverage of robotaxi expansion in Texas documented riders deposited on the opposite side of a freeway from their destination, riders left a fifteen-minute walk away, and rides encountering repeated navigation failures at a single intersection that required a human monitor to clear.¹² Bloomberg found pickup and dropoff behavior unpredictable, with vehicles stopping across the street or several buildings away despite available curb space, and in one instance discharging a rider into an alley.¹³

The prevailing mitigation is to let the rider drag a pin on a map. This presumes app proficiency, a usable data connection, and a rider who already knows the correct coordinate, which is precisely the knowledge they lack. It addresses the symptom while leaving the cause intact: the vehicle’s intelligence is entirely forward-facing. It processes the road ahead in sophisticated detail and the person waiting for it not at all.

The demand here is established rather than speculative. Riders resolve this problem continuously with human drivers by speaking to them. That behavior disappeared only when the driver did.

3. Operator Economics

A coordination failure carries direct cost against a unit economic model already under pressure. A ride that fails to complete forfeits its revenue while retaining its cost. A rider who experiences a failed pickup and returns to conventional ride-hail removes future revenue entirely. The most acute cost, however, is escalation to a human.

Remote assistance is the binding constraint on robotaxi unit economics. Waymo currently runs approximately seventy remote assistance agents on duty at any given time against a fleet of roughly three thousand vehicles, a ratio near one agent per forty-three vehicles that already exceeds what most competitors target as a future aspiration.¹⁴ Industry models place remote operations at $0.05 to $0.15 per mile and project the supervised-vehicles-per-operator ratio improving from roughly six-to-one today toward twenty-six-to-one by 2035, with third-party analysts estimating Waymo’s 2025 cost per revenue mile near $1.98.¹⁵ Every avoidable escalation consumes the scarcest resource in the business and directly degrades the ratio that determines whether autonomy is cheaper than a driver.

These costs concentrate in the final segment of the trip. The driving itself, the portion that consumed a decade of capital and research, largely works. The remaining loss sits in the last fifty feet, where the vehicle and the rider must agree on a location and currently cannot.

4. The Platform: Three Components, One Goal Interface

Curbside deploys as software on the vehicle’s existing onboard compute. It requires no additional sensors, no hardware retrofit, and no cloud dependency in the decision path. It occupies the goal-planning layer directly above the driving stack, consuming perception the vehicle already produces and emitting a destination the driving stack already knows how to accept.

4.1 Referential Grounding

A fine-tuned vision-language model ingests the vehicle’s existing perception feed, including camera streams, bird’s eye view occupancy, tracked objects, and map and curb layers, together with the rider’s dialogue. It outputs a grounded point in the scene and a natural language reply. When a rider states that they are near a blue awning at the rear of a building, the model resolves that reference against what the vehicle can actually see.

The architecture deliberately assigns the model referential reasoning rather than metric estimation. The model identifies which object in the scene the rider means, and the vehicle’s own lidar and localization convert that identification into a metric pose. The model therefore cannot be wrong about distance, because it is never asked to estimate distance. This confines the learned component to the task it performs reliably and leaves geometry to the sensors built to measure it.

4.2 The Viability Gate

Every proposed destination passes through a deterministic verifier before it reaches the driving stack. The verifier evaluates the candidate pose against no-stop zones, curb regulations, and planner reachability. A request that fails any check is never proposed, and the rider receives a spoken explanation rather than a silent failure.

This constitutes planning-level filtering, deliberately distinct from the execution-level safety the driving stack owns. The gate determines whether a request is a sensible thing to attempt. The driving model determines, in the moment, whether the resulting maneuver is safe to perform. Two independent checks at two levels, with the certified stack holding final authority, is a materially stronger safety position than a single agent asserting control.

4.3 The Goal Interface

Verified goals are proposed to the driving stack through a request-and-acknowledge interface. A proposal is treated as real only upon acknowledgment. The stack may accept, defer, or reject it, and the outcome is narrated to the rider rather than assumed.

This interface is the basis of fleet portability. Every autonomous driving system is directed by something upstream. End-to-end architectures map raw sensor input to a plan or control output,¹⁶ yet remain goal-conditioned: state-of-the-art end-to-end planners generate trajectories toward a target supplied to them rather than selecting the destination themselves.¹⁷ Any vehicle operating as a taxi must accept a destination from outside itself. Because that interface necessarily exists in every architecture, Curbside integrates across fleets rather than inside one, negotiating the richest goal representation each stack accepts and degrading gracefully where the interface is coarse.

5. Safety Architecture

The most consequential capability in the system is refusal. A rider who asks to be released in a live traffic lane receives a declination and an alternative: the vehicle identifies a legal pull-over ahead and explains why it will stop there instead. The system neither fails silently nor complies dangerously.

The division of authority is precise and constitutes the substance of the safety case. Curbside determines the target region and intent, meaning which entrance, which side of the building, and what the rider requires. The driving model determines the exact stopping pose within that target and retains complete authority over execution safety. Curbside does not steer, brake, or plan trajectories, and the certified autonomy stack is not modified. A robotaxi already refines a coarse goal into a precise, safe stop. What it does not do is select the correct goal to refine toward when the rider is at the loading dock and the map indicates the front entrance. That upstream decision is the failure point, and it is the only decision Curbside owns.

6. Development Status

The first proof of concept, Curbside 0.5, is complete and is the system shown in our demonstration. It decomposes spatial awareness into discrete detection, depth, and geospatial stages, coordinated by a locally-running fine-tuned vision-language model. Trained and evaluated end-to-end in NVIDIA Isaac Sim, it achieves approximately an 87 percent success rate on randomized simulation tests. Its principal limitation is latency: executing several model stages sequentially on consumer hardware introduces delay that is tolerable in testing but must be reduced before customer deployment.

Curbside 1.0, currently in development, collapses those discrete stages into the single grounded model described in Section 4.1, with the deterministic verifier of Section 4.2 replacing heuristic filtering. Simulation bootstraps supervision, and the 0.5 modules are being repurposed as offline auto-labelers for real logs. Once the system operates on fleet data, every completed pickup self-labels, since the pose at which the vehicle actually stopped serves as ground truth. We are targeting a functional version of the current prototype within weeks, subject to hardware and compute availability.

7. Go-to-Market

Curbside is licensed to fleet operators per vehicle per month. Initial targets are mid-size operators scaling commercial fleets without in-house multimodal research organizations, for whom a rider-interaction layer represents a purchase rather than a build. The initial ask of a design partner is fleet logs and a single pilot route, not a contract.

At $300 per vehicle per month and 30 percent penetration of the United States commercial fleet as it stands today, addressable revenue is approximately $4.3 million annually. At the same price and penetration against Goldman’s projected global fleet of six million vehicles,⁹ that figure is roughly $6.5 billion annually, under two percent of the projected value of the market being served. The transaction is a software procurement decision rather than an infrastructure commitment, which compresses the sales cycle relative to hardware or platform alternatives.

The competitive question concerns scope rather than capability. Large operators can build a version of this internally, and their patent activity around rider identification and drop-off experience indicates active interest.^11, 18 Anything built internally, however, serves one fleet, and the operators most exposed to coordination failure are those without the research capacity to build it at all. The platforms that own the rider relationship face the inverse constraint: they hold rider intent but have no access to vehicle perception, so a rider’s sentence cannot reach the car. Curbside occupies the single position that touches both, across fleets rather than inside one.

8. Closing

The autonomous vehicle industry has solved the harder version of this problem. Driving safely through a complex physical environment is substantially harder than interpreting a sentence about a blue awning. What remains unsolved is smaller, more immediately costly, and structurally invisible to the driving stack, which registers no fault when a rider is left on the wrong side of the street.

The vehicle does not know what the rider knows. It cannot hear what they say, cannot see what they see, and cannot distinguish a reasonable request from a dangerous one. Every ambiguity falling into that gap becomes a failed ride, an operator escalation, or a lost customer. Curbside closes it without modifying the driving stack, without additional hardware, and without a cloud dependency in the decision path. The vehicle already knows how to drive. It does not yet know where you actually want to go, or how to tell you when it cannot take you there.