Robotics News: Wayve's self-driving tech is headed to US cars made by Stellantis — May 26, 2026

Stellantis bets on map-free neural driving, ECARX writes a $750M robotaxi check, and π0.7 keeps emerging.


This was the week the "no HD maps, no LiDAR religion" school of self-driving cashed a real OEM contract — and the week purpose-built robotaxi hardware stopped being a Zoox-only story. Underneath the headlines, foundation models kept inching toward something that survives outside a single robot embodiment. The gap between staged demo and reliable deployment is still wide, but the contracts being signed this week are five-year bets that the gap is closing.


Watch & Listen First


Key Takeaways

  • End-to-end neural driving has its first US OEM contract. Wayve will ship into Stellantis vehicles by 2028 — no HD maps, no sensor lock-in.
  • Robotaxi economics are getting rebuilt from the chassis up. ECARX and May Mobility's $750M framework targets a 50% hardware cost cut by skipping retrofits entirely.
  • Foundation models are starting to generalize across embodiments, not just tasks — Physical Intelligence's π0.7 and NVIDIA's GR00T N1.7 both posted real cross-platform numbers this month.
  • Humanoid deployments outpaced humanoid demos. JAL at Haneda, Agility at Toyota Canada, and BMW Leipzig are now multi-unit operational, not pilots.
  • Counter-drone is the new growth budget. A $500M ceiling for Perennial Autonomy signals the Pentagon spending where the threat actually is.

The Big Story

Wayve's self-driving tech is headed to US cars made by Stellantis · May 21, 2026 · TechCrunch
Wayve's pitch has always been the bitter-lesson one: an end-to-end neural net that consumes whatever cameras and radar the OEM already ships, with no HD map dependency. Getting a prototype running across Stellantis' platforms in under two months is the technically interesting claim — it implies the model genuinely is sensor-agnostic, not paper-agnostic. The 2028 production date is generous, but this is the first time map-free, learning-based driving has a binding consumer-vehicle contract in the US, and Mobileye / Tesla now have a same-stack competitor with someone else's brand on the hood.


Also This Week

ECARX and May Mobility sign $750M deal for purpose-built robotaxis · May 19, 2026 · TNW
Building L4 sensor and compute integration at the factory line — not bolting it onto a Pacifica — is the only path to the cost curve robotaxis actually need; the offshore manufacturing structure to clear US ICTS rules is the playbook other Chinese-backed AV suppliers will copy.

Pentagon backs AI counter-drone startup Perennial Autonomy with $500M ceiling deal · May 21, 2026 · DroneLife
Drone-on-drone autonomy is now a discrete DoD program of record, and Merops / Bumblebee / Hornet interceptors crossing the threshold from research to JIATF-401 contract validates that swarm-defeating swarms have moved past PowerPoint.

Northrop YFQ-48A Talon Blue completes autonomous taxi tests · May 18, 2026 · Army Recognition
Independent runway handling is the boring milestone that gates the interesting one — the CCA program is now days away from its first uncrewed autonomous takeoff, and Northrop is ahead of Anduril on the public timeline.

Romark Logistics deploys Dexory autonomous inventory robots · May 19, 2026 · Robotics & Automation News
Mobile scanning robots that don't disrupt picking throughput are the easy-to-deploy wedge that finally makes "digital twin of the warehouse" a real, live data layer rather than a quarterly stocktake.

Hexagon and Fill Maschinenbau deploy AEON humanoid in Austrian factory · Week of May 19, 2026 · Interesting Engineering
Machine tending and inspection is the unsexy beachhead — and Hexagon's metrology stack underneath AEON gives this deployment a closed perception-feedback loop that BMW Leipzig has been quietly proving works.


From the Lab

SOMA: Spatial Memory for Out-of-Vision Manipulation in Vision-Language-Action Models · arXiv 2505
VLAs get brittle the moment the target object leaves the camera frame — SOMA bolts a persistent multi-view spatial memory onto the policy, cutting target re-localization time 40–59% and head-search path 40–57% across five real-world tasks. It's a small architectural addition that addresses one of the most-cited failure modes in production-grade π0 and GR00T deployments, and it's the kind of fix that quietly closes the gap to "you can leave the robot alone in a real kitchen."


Worth Reading


The week the AI of self-driving stopped being a Tesla-only argument, and the week we stopped pretending purpose-built robotaxi chassis were optional.