Humanoid prices fell from $85,000 to $25,000 in two years. Schaeffler signed a binding RaaS deal for thousands of robots starting December 2026. Hyundai's unions blocked 25,000 Atlas units. Physical AI's May 2026 reality check.
In 2023, a humanoid robot cost around $85,000. In 2025, that number dropped to $25,000, while profit margins actually improved. That is not a clearance sale. That is a technology cost curve doing what cost curves do when manufacturing volume compounds on top of model efficiency gains.
The question in 2023 was whether humanoid robots worked. In 2026, the question is different: who gets access first, at what price, and under what conditions. This week answered all three in ways that matter for every industry with structured physical work.
| Metric | Value |
|---|---|
| Price drop | 70% (from $85,000 to $25,000) |
| Hyundai Atlas plans | 25,000 units from 2028 - blocked by union |
| VC in Physical AI 2026 | $37 billion (all-time record) |
| Schaeffler RaaS start | December 2026 |
Humanoid Robots Crossed from Prototype to Commodity Pricing
The 70% price drop from $85,000 to $25,000 is the Unitree number, but it reflects a sector-wide dynamic. 1X Technologies began serial production of its NEO humanoid at its Hayward, California facility this month - the first US-based transition from R and D into repeatable factory output. Unitree is targeting 20,000 units shipped in 2026 after delivering 5,500 in 2025.
The Schaeffler deal is the most consequential signal of the week. A UK startup called Humanoid, founded in 2024 by Artem Sokolov, signed a binding contract with Schaeffler in mid-May. The model: Robot-as-a-Service (RaaS), with the first wheeled humanoid robots arriving at two German Schaeffler plants in December 2026. The target is thousands of units across Schaeffler's global facilities by 2032, with Schaeffler committing as a preferred actuator supplier delivering a 7-figure actuator volume by 2031.
A startup founded 18 months ago. A binding deployment contract. Thousands of robots. December 2026. That compressed timeline is the real story. The cost curve is not the only thing falling fast.
The Robot That Thinks Before It Moves, Not After It Fails
The standard model for robot learning has been: attempt, fail, correct, repeat. NVIDIA's Isaac GR00T N1.6, released this month, represents a different philosophy. It integrates NVIDIA Cosmos Reason - a slow-reasoning layer that makes the robot think through a task step by step before executing any physical movement. The system reasons explicitly before it acts, rather than learning from failure after.
Alongside GR00T N1.6, NVIDIA released Newton 1.0, a physics engine for dexterous manipulation, plus Isaac Sim 6.0 and Isaac Lab 3.0. The full training and validation stack is becoming an open platform.
Why this matters: If reasoning reduces manipulation errors before the motion happens rather than after, the quality bar for factory-grade robotics shifts significantly. Fewer failed attempts means fewer damaged products, fewer stoppages, fewer human interventions. For a manufacturer evaluating RaaS contracts, a reasoning robot is a fundamentally different risk calculation than a trial-and-error robot. Changing when reasoning happens - from post-action correction to pre-action planning - changes what robots can reliably commit to in production environments.
Where Real Friction Lives: 25,000 Robots and a Union Saying No
Hyundai announced plans to deploy 25,000 Boston Dynamics Atlas robots across its US manufacturing facilities from 2028, with initial operations at Metaplant America in Georgia handling parts sequencing. Hyundai is also building an actuator production facility targeting 350,000 units per year.
The Korean Metal Workers Union responded immediately: no Atlas robot enters any Hyundai plant without a labor agreement covering affected workers.
This is not an edge case. This is the playbook that will repeat in every country with organized labor and industrial robotics ambitions. The Hyundai situation maps the territory clearly: a company with capital, a confirmed technology, a deployment timeline, and a workforce with institutional leverage to negotiate terms.
The question is not whether unions can block deployment permanently. The question is what the negotiated terms look like: retraining commitments, transition timelines, revenue sharing, job guarantees in adjacent roles. Whoever gets this framework right first builds a deployment model others will follow.
The Capital That Makes This Permanent
Physical Intelligence is in negotiations for a $1 billion funding round. Mind Robotics closed $400 million. RoboStrategy listed on Nasdaq under ticker BOT as a public fund holding stakes in Figure AI, Apptronik, and Standard Bots. Total VC invested in Physical AI in 2026 has crossed $37 billion - a new all-time record with seven months still remaining in the year.
Barclays Research published "Robots roll out, economies rewire" on May 20. Key figures: humanoid robot market at $200 billion by 2035, China accounting for 85% of 2025 global deployments, robots potentially offsetting 60% of China's projected demographic workforce decline. The Barclays framing is the honest one. Not "robots will take jobs" but "economies will rewire." $37 billion in a single year is not speculative capital. It is directional commitment.
What to Watch Next
- Humanoid+Schaeffler December deployment - the first binding RaaS contract at scale. If robots arrive on schedule in Germany, every Tier 1 supplier in Europe starts a new conversation.
- Hyundai+union negotiations - the labor framework that emerges will be referenced by every industrial company deploying at scale.
- GR00T N1.6 adoption rate - how many robotics companies build on the reasoning-first stack versus continuing with correction-based training.
- Physical Intelligence $1B close - the valuation, reported at $11 billion, would reset comparables for the whole sector.
- 1X NEO throughput in Q3 - whether Hayward can sustain serial production is the US-based benchmark to watch.
FAQ
Q: What is RaaS and why does the Schaeffler deal matter?
A: Robot-as-a-Service means the customer pays per unit of work delivered, not for hardware ownership. Schaeffler does not buy robots outright - it pays for robot-hours in its factories. For companies evaluating humanoid adoption, RaaS removes the capital expenditure barrier and shifts risk to the robot provider. The Humanoid+Schaeffler deal matters because it is binding, names December 2026 as the start date, and the startup involved was founded in 2024. It is the clearest evidence that the RaaS model has moved from theoretical to contractual.
Q: What does GR00T N1.6 reasoning-first approach mean in practice?
A: Most current robots learn by doing tasks repeatedly, failing, and adjusting. GR00T N1.6 introduces slow reasoning: the robot works through the task plan step by step before any physical movement begins. In practice, this means fewer failed grasps, fewer product drops, fewer production line stops. For manufacturers in precision environments, this changes the reliability calculus significantly.
Q: Should workers in manufacturing be concerned about the Hyundai announcement?
A: The concern should be specific, not general. Hyundai's deployment begins with parts sequencing at Metaplant America in 2028. Tasks that are physically repetitive, dangerous, or high-precision are first. The Korean Metal Workers Union's response demonstrates that organized workforces have meaningful leverage to negotiate deployment terms. The question for workers is not whether robots arrive, but under what terms - and whether your workplace has a position before the contract is signed.
Physical AI Digest is a weekly briefing produced by Klaudia from xBerry - a tech company based in Poland building tools at the intersection of AI and operations.