1. EXECUTIVE BRIEF

▪  State Grid Corporation of China budgeted ~6.8 billion yuan to procure roughly 8,500 robots — about 5,000 robot dogs plus humanoids and dual-arm systems — for 2026 power-grid operations across 600+ defined task categories, with Unitree, Deep Robotics, AgiBot, UBTech and Fourier Intelligence named as suppliers; the largest single Physical AI procurement publicly disclosed at this scale.

▪  Skild AI absorbed Zebra Technologies' Robotics Automation business (formerly Fetch) on 15 April with Zebra taking equity in Skild; Pudu Robotics closed a ~$150M Series B2 at over $1.5B on 23 April; Reliable Robotics raised $160M with 200+ commercial and military commitments and a 2026 operational start — deployment-funded rounds led the week's capital flow.

▪  Apptronik appointed former Waymo CPO Daniel Chu as Chief Product Officer on 28 April to push Apollo from prototype to production scale, the same week Path Robotics, XPeng and others continued to publicly mark the 90% → 99% reliability gap as the binding constraint on humanoid economics.

2. DEEP DIVE — THE BUYER FLOOR

Last week Erlangen showed what one humanoid can do for eight hours. This week, the world's largest electric utility by customers served signed off on enough robots to make eight thousand of those questions matter at once.

What happened.

State Grid Corporation of China — serving roughly 1.1 billion customers across 26 of 31 mainland provincial-level regions — has allocated approximately 6.8 billion yuan for 2026 procurement of embodied-intelligence systems. The plan covers around 8,500 units: roughly 5,000 robot dogs for substation patrol and inspection across mountainous terrain, remote sites and hazardous corridors, plus humanoids and dual-arm robots for ultra-high-voltage maintenance work. The procurement spans more than 600 defined task categories. Five Chinese OEMs are named as suppliers: Unitree, Deep Robotics, AgiBot, UBTech and Fourier Intelligence. China Southern Power Grid and other utilities lift the 2026 sector total above 10 billion yuan. The original disclosure came from Chinese-language outlet Jiemian, with SCMP corroboration. Treat the 8,500-unit headline as an authorisation to spend, not yet a delivery number.

Why operators care.

Western utilities — PG&E, EDF, E.ON, National Grid — have published nothing comparable. The reliability bar that humbles humanoids on a Siemens electronics line, where 99% is the floor, does not bind in the same way for substation patrol, where the alternative is sending a person up a tower in a thunderstorm. State Grid is buying for tasks where current capability is already the better option, not the prestige application. That changes the deployment calculus: pilot-to-production conversion stops being the bottleneck when the buyer pre-commits to thousands of production seats.

Why investors care.

Captive demand floor. Apptronik at a reported ~$5.5B post-Series A-X, Figure at a reported $39B, Unitree at a target $7B Shanghai STAR valuation — every Western and Western-priced humanoid OEM is being valued against a buyer landscape that has not yet committed at this scale. State Grid's commitment is procurement, not pilot. The lifetime value of being on a 26-province utility's supplier list with thousands of confirmed orders is structurally different from the LTV of a company still chasing letters of intent.

What could go wrong.

Reliability incidents on public-facing Chinese deployments are now near-weekly — a Unitree humanoid reportedly malfunctioned at a festival this week, swinging aggressively before shutdown. State Grid's 600 task categories are not all sub-99% acceptable; ultra-high-voltage maintenance at altitude has unforgiving failure consequences. Authorisation-to-spend converts to delivered, deployed, working hardware on a curve, not a cliff. The original number rests on single-source Chinese reporting; the absolute count and timeline should be treated as direction, not invariant. The bear case from Issue #06 has not weakened — it now has a much larger pool of robots to break in public.

3. CAPITAL LAYER

Skild AI × Zebra Robotics Automation — acquisition closed 15 April; Zebra takes equity stake in Skild as part of the deal. — Foundation model buys distribution.

Skild ($14B post Series C in January) absorbs Zebra's Robotics Automation business — the same week Skild's brain went live on the NVIDIA-Foxconn Blackwell GPU assembly line in Houston. The horizontal-platform thesis acquires a deployed warehouse install base.

Pudu Robotics — Series B2, ~$150M, 23 April. Values the Shenzhen-based service-and-industrial robotics maker at over $1.5 billion. — China's robotics capital stack now extends past humanoids.

PRNewswire primary; The Robot Report and Tech in Asia confirmed. Pudu has shifted from hospitality service robots into industrial logistics deployments. Series B2 is the deployment-funded round, not the demo round.

Reliable Robotics — $160M, 21 April, led by Nimble Partners. 200+ commercial and military orders for autonomous cargo aircraft pursuing FAA type certification; 2026 operational start. — Capital, certification path, named customers and ops timeline in one round.

The rare combination in regulated autonomy. The cargo-aviation autonomy lane is now financed for delivery, not proof.

Stereotaxis × Robocath — $20M upfront + $25M milestones, 14 April. Endovascular surgical robotics consolidation; milestone payments tied to FDA clearance, mid-2026 close expected. — Surgical-robotics M&A heats.

One platform with combined cardiac-and-vascular reach against two point solutions.

4. ENGINEERING FLOOR

Apptronik appoints Dan Chu as CPO (28 April). Former Waymo CPO and former 23andMe CPO. The hire follows February's $520M Series A-X that brought total Series A funding to $935M at a reported ~$5.5B valuation. — The execution layer hires its product discipline. Apollo is past prototype; Chu is hired to run production scale-up. Pulling AV-grade product leadership into a humanoid OEM is the same translation Reliable Robotics is making on the regulated-autonomy side. The pattern: capital is being deployed against operating-team upgrades, not platform pivots.

TI / NVIDIA / Lattice radar-camera fusion stack (28 April). Reference architecture combining TI mmWave radar, Lattice sensor bridge, and NVIDIA Holoscan for low-latency humanoid perception. — Multi-vendor perception integration template. Moves the perception layer from research code to deployment-ready edge stack. The radar layer is what makes a humanoid usable in low-light, dust and partial-occlusion environments — exactly the State Grid substation context.

Skild Brain on the NVIDIA-Foxconn Blackwell line. First public mass deployment of Skild's omni-bodied foundation model — the production line for Blackwell GPU server systems in Houston. — The brain layer ships into a real factory the same week the demand layer publishes its budget. Foundation-model claims convert to operational data only when robots run in production.

Rapyuta PA-AMR field metrics (MODEX, 13–16 April). 50% reduction in worker walking distance, 2x picking productivity, 30% search-time cuts at deployed warehouse sites. — Operational metrics from running deployments — the baseline against which humanoid economics will be measured. PA-AMRs are working today; humanoids are catching up to a moving target.

Policy & Regulatory Radar — NHTSA FMVSS No. 110 amendment for ADS vehicles. Federal rulemaking proposed 1 April; comment period closed 1 May. — Operators with positions on the rule should now watch the agency's response and any companion rulemaking on Part 555 exemptions.

5. CHINA DESK

State Grid procurement — supply-chain reading. The 8,500-unit commitment names five OEMs (Unitree, Deep Robotics, AgiBot, UBTech, Fourier Intelligence). Unitree's grid-services revenue exposure now lands alongside its Shanghai STAR IPO process — public-market investors are about to see embodied-intelligence revenue priced as a utility-services line, not a research-and-consumer line. Original via Jiemian, Chinese-language; SCMP cross-reference.

UBTECH × Airbus — first humanoid deployment in aviation manufacturing. Fresh extension of UBTECH's industrial customer roster after BYD, FAW Group, Foxconn, Audi and Siemens. Aviation manufacturing tolerances are tighter than automotive — this is a reliability statement aimed at industrial buyers who care about it.

Sector demand pool exceeds 10B yuan. State Grid is not alone. China Southern Power Grid and other utilities push 2026 embodied-intelligence procurement above 10 billion yuan. Captive demand at this scale — across grid, defence and state-aviation procurement — is what makes the Chinese OEM cost curve durable, even when individual robots underperform.

6. GLOBAL PULSE

Korea — Hyundai targets 30,000 Atlas humanoids/year by 2028 at HMGMA Savannah. Process-by-process validation: parts sequencing 2028, component assembly 2030. 2026 fleets already committed to Hyundai and Google DeepMind. — Korea's national champion publishes manufacturing scale that exceeds most US peers.

Canada — Toyota TMMC adds Agility Digit at Woodstock for RAV4 logistics. 7+ commercial Digit units active under Robot-as-a-Service pricing. — Tariff-pressured North American auto production responds with humanoid automation, not relocation.

EU — UN-endorsed DCAS rules expand hands-off Level 2+ overtaking; EU type-approval framework progresses through 2026. ETSC has warned Europe lacks NHTSA-style crash reporting infrastructure. — Capability-permitted before reporting-required is a recurring EU pattern in autonomy regulation.

7. RESEARCH ROOM

π0.7: a Steerable Generalist Robotic Foundation Model with Emergent Capabilities — Physical Intelligence et al., arXiv:2604.15483, posted 16 April 2026 (V2 24 April). The paper introduces π0.7, a 4B-parameter VLA built on Gemma3 with an 860M-parameter action expert; the central claim is compositional generalisation — the model recombines skills from different training contexts to attempt novel tasks (e.g. operating an unfamiliar air fryer with language coaching, success rate rising from ~5% to ~95% in roughly 30 minutes of prompting). For State Grid's 600 task categories, the relevance is direct: the binding question is no longer whether a robot can do one task at 99%, but whether one model can be steered through hundreds at acceptable reliability without per-task retraining.

8. OPERATIONAL TAKEAWAY

The week's signals fit. Issue #06 closed with the supply side publishing a benchmark — one humanoid, eight hours, 90% on a Siemens electronics line. Issue #07 opens with the demand side publishing a budget — one utility, ~6.8 billion yuan, 8,500 units across 600 task categories. Between them, the execution layer caught up: Apptronik bought AV-grade product leadership; Skild's brain shipped to a working assembly line; Pudu and Reliable Robotics raised on backlog rather than promise; π0.7 published a compositional-generalisation result that is the closest thing to a Western answer to “can one model run 600 task categories.” The bear case wrote itself in parallel — Wolf's babysitting-ratio analysis, the Unitree festival incident, the persistent 90%-to-99% gap. But the operator question shifted. It is no longer “will the customer come.” It is “is your portfolio building toward the customer that has.”