# Vehicles become intelligent robotic platforms that autonomously perform real-world tasks beyond driving.

Topic radar · 726 documents · 2021-01-01–2026-10-21 · updated 2026-10-21

Canonical: https://slate.greyb.com/radar/vehicles-become-intelligent-robotic-platforms-that-autonomously-perform-real-world-tasks-beyond-driving/innovations

## In one minute

### A machine learns a task by watching an operator do it
The most complete build in this material belongs to AIM, and it is a closed loop rather than a model. A data-gathering module runs on the machine while a human operates it, recording control inputs, vehicle state and sensor streams — GPS, inclinometers, hydraulic pressure — alongside LiDAR and camera data that keep a three-dimensional site map current. That record goes to a remote machine, where a model predictive control model of the vehicle's behaviour and its surroundings is trained, and the trained model is pushed back down to the vehicle to run the task on its own [1] [2].

The detail that makes it more than a description is how the demonstrations are pooled. To train one model on several machines of the same type, the filings translate each run's poses onto a common starting pose before training, so the model learns the machine type rather than the individual unit and a fleet becomes the training set [3]. The hybrid is deliberate: inverse kinematics for the tool comes from a decision-tree model and the hydraulics are simulated from physics, so the learned part is bounded by a known model of the linkage [2]. A Hong Kong construction-robotics group files the same idea from the other direction, an end-to-end multimodal imitation learner that takes instruction, machine status and terrain and emits control commands, driven from a handset across several machines at once [4].

1. [US2026286651A1](/?paper=US2026286651A1&utm_source=radar&utm_medium=dashboard)
2. [WO2026143227A1](/?paper=WO2026143227A1&utm_source=radar&utm_medium=dashboard)
3. [US2026286653A1](/?paper=US2026286653A1&utm_source=radar&utm_medium=dashboard)
4. [CN121742281A](/?paper=CN121742281A&utm_source=radar&utm_medium=dashboard)

### What is on show is a release, not an order
Faraday Future scheduled its first robot delivery for 27 February 2026, to a Florida vacation-rental operator, and by April was publishing video of a quadruped completing a food delivery with no human intervention. Robot.com opened for business on a tally — 1.7 million completed tasks — with Sodexo and AWS named as production customers. XPeng showed its Iron humanoid at Auto China running the same vision-language-action model as its driver assistance, with first deployments billed for controlled service environments.


### The commitments are to compute and to test rigs, not to machines
Read the financings together and the buyer of last resort is not a fleet operator. NVIDIA's venture arm and the UK's National Wealth Fund put 103m into Oxa's industrial autonomy software in the same round, 50m of it public money. Chinese vehicle silicon took $200m into SiEngine, whose chips pair with ECARX central computing platforms under a multi-year supply agreement with a global automaker. Honda is investing in and co-developing a vehicle SoC with Mythic directly. The pattern is a commitment to the substrate a machine runs on, funded by the parties that sell that substrate or want the capability domestically.


### The implement is part of the control system
AIM's steering correction does not work through the wheels at all. When the actual heading diverges from the target path by a threshold, the machine lowers one side of the blade into the ground to pull itself back onto line, with the depth of the cut set by reinforcement learning, imitation learning or a simple exponential function. That is worth stating plainly: on an earth-moving machine the tool is an actuator, and the autonomy layer has to model how digging changes the vehicle's path, not just where the vehicle is pointing. It also explains why the retrofit filings and the task filings converge — if the lever and the blade are both under software control, the difference between a converted machine and a purpose-built one narrows.


### The order that repeats
FairPrice's sequence is the clearest commercial path in this material. A cargo trial with Zelos in October 2024, approval from Singapore's transport authority to move to remote operation, then an agreement to add about 30 Zelos Z10 vehicles on a route just under two kilometres between two distribution centres.


### The robotics bet is being financed outside the car company
XPeng raised more than 900m for its robotics unit at a valuation above 6.3bn, a record for a single private round in Chinese embodied AI. Hyundai committed to 25,000 Atlas units in its US plants, 30,000 a year of manufacturing capacity and roughly 300,000 actuators a year, alongside supplying Waymo with robotaxis. Around twenty Chinese automakers have entered robotics by subsidiary, R&D line or investment. None of these reads like a product line; they read like options held on a separate balance sheet, which is the strategically correct shape if the machine business turns out to be a component business rather than a vehicle business.


### Two ways to reach the controls, and the trade between them
The mechanical route puts pistons on the levers and pedals and drives them from a microcontroller, which works on any machine regardless of who built its electronics. The electronic route bolts onto the bus instead: Niteko's unit interfaces with existing control units, explicitly avoids the safety circuits and any invasive mechanical change, runs on a single-board computer over industrial CAN, and stores and replays routes without vision. The trade is concrete. Actuators reach any machine and yield no state feedback; the bus yields state and faults but only from a machine that exposes them.


### The first paying customer is a programme with a budget
Aurrigo's entry into north-east England is a £0.9m project with the Tees Valley Combined Authority, and its Changi work runs under a multi-year partnership partly funded by Singapore's civil aviation development fund. The Port of Tyne demonstration was delivered through the UK government's CAM Pathfinder programme. On the military side, American Rheinmetall holds an eighteen-month contract under the US Army's Project Sustainment, built on Harbinger's hybrid vehicle.


## Overview


Read end to end, the field is building two businesses and describing a third. One business fits an autonomy layer to a machine that already works: a dozer, a terminal tractor, a tractor, a side-by-side. The other carries passengers on roads. The third — a road vehicle that also performs a physical task — is the one the framing names, and it appears in filings and in nobody's operating record.


## Research

### Safety is enforced as an envelope, not as a rulebook
Rather than argue that the task was performed correctly, these machines are given physical limits and are pulled back inside them. AIM applies the brake when LiDAR-derived slope crosses a threshold. A Guangxi University filing fuses LiDAR, camera, GPS, IMU, wheel-speed and yaw data through an extended Kalman filter, predicts roll angle and load transfer rate with a trained network, and re-plans or detours the path when load transfer exceeds a maximum. Sumitomo takes the geometric version: two machines work adjacent areas, but each machine's permitted range is set back from the shared boundary so the conflict has nowhere to occur.


### Fleet coordination is resource allocation, and the learned part is a scoring term
The scheduling filings describe the same architecture repeatedly: a central server owns the global problem and the vehicles own the local one. Underground loaders get a discretized spatiotemporal model of the roadway topology, with resources pre-occupied so two machines cannot claim the same segment at the same moment, and onboard terminals free to make local decisions when the server's instruction does not resolve. Swarms get a generalized task model that decomposes formation-level objectives down to platform-level control, with replanning triggered by situational updates. An air-ground inspection system grows its raster map as the vehicles explore, so mapping and task execution become the same act.


### Who is allowed to make the machine move
One filing treats actuation as a security boundary rather than a control problem. A robot executes a server-signed trip plan, authenticates the recipient on arrival through a short code or biometrics, and only then does a hardware gate energise the actuators; after the task it cryptographically erases transient data and uploads a signed audit log. It is a small delivery robot, but the mechanism is the one a work machine on a shared site would need, and it sits below the network layer where most of the discussion sits.


### What autonomy actually returns where it has been measured
The field numbers are modest and specific, and they are the best evidence in this material that the layer pays. An autonomous agricultural vehicle ran 127 seeding, spraying and spreading operations over more than 5,700 hectares across three years, beating conventional equipment on seeding efficiency and fuel use and matching it on spraying and spreading. A commercialised orchard sprayer, characterised after a valve fault was fixed, reached 91% left-right symmetry but only 57% spray uniformity across the spray height — good enough to sell, not good enough to stop engineering. An AGV carrying an X-ray source for aircraft wing inspection improved handling and posture-adjustment efficiency by over 70% per aircraft and removed one operator from the job. A weeding implement detects blockage from strain on its shank and overrides its own actuator before the tine jams.


### Validation is the binding constraint, and the twins are graded
A maturity assessment of vehicle digital twins finds most of them at the monitoring level, with the twins that permit restricted autonomous operation existing only in closed environments. That is the same perimeter the deployments describe, reached from the software side: the places that can be simulated to the required fidelity are the places already fenced.


### The layer has a price, and the deployments have a perimeter
The autonomy stack is now a bill of materials. Bot Auto puts compute and sensors for a Class 8 truck at about 50,000 using an Nvidia Hyperion-based domain controller, against 150,000 to $400,000 implied for other systems, and reports roughly 350 loads a month across Texas — a vendor figure set against a third-party cost guide, not an audited industry price.


## Market

### The chassis makers are walking onto the same ground
Oshkosh used a US Army trade show to present three production-ready variants of a multi-mission autonomous vehicle family, pitched as payload-agnostic platforms that are ready now. That is the vehicle manufacturer selling autonomy as a catalogue option, competing for the same procurement as the software vendors who fit a stack to whatever chassis the customer already owns.


### Where the working vehicle already has a shelf
IDRIVERPLUS's agreement with Henan Senyuan pairs an autonomy supplier with a specialty-vehicle manufacturer whose lines are firefighting, emergency response and sanitation vehicles, with the supplier claiming the leading share in several deployment scenarios. That is the market structure in one deal. The machine already has a procurement category, a bodybuilder and a municipal buyer, and the autonomy arrives as an option fitted to it.


### A programme that stopped, and where it puts the buyer
FedEx ended its Roxo delivery robot in 2022, saying the machine did not meet the near-term value requirements of its cost programme. The carrier kept the business and dropped the robot.


## Strategy

### The durable IP is being placed in the tool and the schedule
Where a company has built a connected family rather than a single application, it is AIM, and every one of those filings is about making an earth-moving machine work: training from an operator's own runs, pooling demonstrations across a machine type, steering by lowering the blade, enforcing a slope limit through the brake. Around that, the filing record is single applications from universities, state metals and mining groups, and port terminals on dispatch and conflict-free routing. The carmakers and robotaxi operators that dominate the announcements are largely absent from it. Effort is being committed by the people who own the work and the site, and by the vendors selling to them.


### The purchase orders that exist are public and small
Overland AI holds $24.5m to field autonomous ground vehicles for the US Army's 75th Ranger Regiment, having started at $20m, and $19.7m from the Marine Corps for more than a dozen vehicles due by early 2027. These are the clearest cases of a customer committing money to a machine that does a task, and they are counted in tens of millions with delivery dates attached. Set that against 25,000 humanoids in a single internal deployment and the asymmetry is stark: the volume commitments are intra-company, and the external ones are contracts a defence programme can sign.


### Validation is being built as public infrastructure, which decides who sets the bar
A national defence programme is funding a digital twin of a ten-kilometre proving ground with real surface undulation, soil behaviour and simulated weather, wired to actual vehicle controllers, running through April 2028. This is the commitment with the longest tail. Once a state owns the simulator that defines robustness, the acceptance criteria stop being a commercial negotiation and become a specification, and every smaller buyer downstream inherits that regime instead of writing its own.


## Areas

- Enabling Software & AI · 206 documents
- Adjacent Robotic Platforms · 160 documents
- Regulation, Safety & Governance · 151 documents
- Core Autonomous Driving & Robotaxi · 101 documents
- Market Adoption & Social Acceptance · 102 documents