Beijing drew the cameras. Two factories earned the week.

While more than 300 robots from 26 brands ran Yizhuang's 21-kilometer loop on Sunday, humanoids in Nanchang and Woodstock, Ontario did something the race could not: they held live production positions on revenue-generating lines.

What happened. On April 14, AGIBOT switched on four G2 humanoids inside Longcheer Technology's tablet manufacturing workshop in Nanchang, Jiangxi. The robots work Multimedia Integrated Testing (MMIT) stations — picking devices off conveyors, placing them in test fixtures, sorting pass and reject units. Throughput is 310 units per hour. Cycle time runs 18 to 20 seconds. Success rate exceeds 99.9% over 140 cumulative operating hours. Downtime stays below 4%. Line integration took 36 hours. AGIBOT plans to scale the site to 100 units by end of Q3.

Six thousand miles away, Agility Robotics went live with seven Digit humanoids at Toyota Motor Manufacturing Canada's RAV4 facility in Woodstock, Ontario, in early April. Digit carries 35 pounds of payload, runs on NVIDIA Jetson AGX Thor, and moves auto-parts totes off automated tuggers. The contract is structured as Robots-as-a-Service — Agility owns the hardware and maintenance burden; Toyota pays for uptime.

Why it matters for operators. Procurement teams evaluating humanoids have spent eighteen months asking two questions: will these machines hold a position across a shift, and what is the integration cost? Nanchang and Woodstock answer both in different templates. AGIBOT delivers a 36-hour integration and a 99.9% task-level success rate on a single-task precision station. Agility delivers RaaS economics that push hardware risk to the vendor. The procurement rubric now has real reference data — not vendor demos.

Why it matters for investors. AGIBOT rolled out its 10,000th humanoid in March 2026. Per Omdia, the Shanghai-based firm took 39% of the 2025 global humanoid market with more than 5,100 units shipped. The 100-unit Longcheer target by Q3 is the first unit-level commitment from a named anchor customer to any Chinese humanoid OEM — the shape of a revenue curve, not a pilot. Agility's RaaS structure at Toyota is the Western mirror: vendor balance-sheet risk traded for recurring revenue and direct deployment telemetry.

What could go wrong. Both deployments are single-task. MMIT station work is repetitive precision, not dexterous manipulation in free-form environments. The 140-hour figure at Longcheer is a starting point, not a reliability proof. Sunday's race surfaced the other side of the coin: Unitree's H1 set a proportional world-record-equivalent pace of 4:13 over 1.9 kilometers in Thursday's qualifier, then ran a competitive ~51:40 split in Sunday's main race before a late-race posture failure ended with the robot stretchered off the course. The gap between works on a line for 140 hours and holds up across 24/7 multi-task deployment for a year is still the commercial question nobody has answered.

Tokens decide what to pick. Torque delivers it to the testing fixture. This week, both halves of the stack drew paychecks.

1X Technologies × EQT — 10,000 Neo humanoids, pilot phase entering 2026 — Largest commercial humanoid procurement to date $20,000 per unit or $499/month subscription. Rollout through 2030 across 300+ EQT portfolio companies — manufacturing, facilities, healthcare. Deal originally struck December 2025; pilot phase now confirmed alongside first Neo deliveries.

Stereotaxis × Robocath — $20M upfront + up to $25M earnouts — Surgical robotics consolidation continues SEC 8-K filed April 15. Earnouts tied to FDA milestone + two commercial milestones; interim €1.3M bridge funds Robocath operations pre-close. Target close July 2026. The earnout structure signals regulatory pathway clarity remains the primary value trigger in medical robotics M&A.

ABB Robotics — Q2 2026 stock listing confirmed at AGM — First pure-play industrial robot public comp since FANUC 100% spin-off approved; shares distributed as dividend-in-kind to ABB shareholders. Likely listings: Switzerland + Sweden. Standalone capital structure gives investors an industrial-robot OEM position without ABB conglomerate exposure.

Held this week — Galaxea Dynamics $280M B+ extension and LingXin SmartHand $2.8B valuation — both [UNVERIFIED]. Only single secondary sources surfaced. Physical AI Weekly will not publish dollar figures without primary confirmation.

NVIDIA Isaac GR00T N1.7 released to Early Access with Apache 2.0 commercial licensing (April 17). The 3-billion-parameter vision-language-action model ships with a Cosmos-Reason2-2B / Qwen3-VL backbone, pretrained on 20,000 hours of EgoScale human video data. The commercial-licensed open VLA path is now available for humanoid OEMs building on Jetson Thor; next-generation GR00T N2, previewed at GTC, is slated for end of year.

First Humanoid Foundation Model Benchmark launches with 40 models rated across 10 capability dimensions (April 14). Dimensions include locomotion, whole-body control, bimanual and dexterous manipulation, navigation, sim-to-real transfer, cross-embodiment generalization, real-time inference, and long-horizon planning. Organizations rated include NVIDIA, Google DeepMind, Tesla, Physical Intelligence, AgiBot, Skild AI, Figure, 1X, Unitree, and nine others. First vendor-neutral reference for operators selecting AI brains across fleets.

"Switch" framework achieves 100% skill-switching success rate on Unitree G1 (arXiv:2604.14834, April 16). The hierarchical system transitions between agile locomotion skills — squats, back kicks, rapid recoveries — at 100% SSR across Easy, Medium, and Hard difficulty levels versus a 2% baseline. Validated on hardware currently shipping to customers. Multi-skill autonomy is moving from lab demo to deployable capability on platforms already in the market.

Policy & Regulatory Radar. No material EU Machinery Regulation or US safety-standard updates affecting humanoid deployment cleared in the week window.

AGIBOT's scale advantage came into full view. Beyond the Longcheer go-live, 2025 revenue hit RMB 1.05 billion, the Q2 2026 order book is sold out, and Omdia confirms 39% global humanoid market share on 5,100+ units shipped. The 100-unit Longcheer target by Q3 is the first named anchor-customer commitment at unit volume for any Chinese humanoid OEM.

Changan Tianshu Intelligent Robot formally capitalized at RMB 450 million (April 1). Chongqing-based; Changan holds 60%. First in-vehicle component robot targeted Q1 2026; humanoid mass production 2028. Changan becomes the fifth Chinese automaker — after BYD, XPeng, Xiaomi, and NIO — to spin out a robotics subsidiary. Auto-to-humanoid is now a structural feature of China's industrial policy, not isolated bets.

UBTECH subsidiary UQI signs Honda Trading partnership on Walker series humanoids and unmanned delivery vehicles. Japanese industrial distribution channels are opening for Chinese humanoid OEMs — a go-to-market vector Western vendors have not matched.

Dubai — Joby Aviation's DXB vertiport goes fully operational. All four Dubai vertiports now complete, with rapid-charging infrastructure live. Commercial service target: by end of 2026. Infrastructure is no longer the constraint — certification and paid-passenger rollout are.

United States — Waymo expanding to four new cities (Baltimore, Philadelphia, St. Louis, Pittsburgh) with the Arizona factory with Magna doubling output to 2,000+ vehicles by end 2026. Manufacturing is now the binding constraint for fleet scale, not regulation or demand.

Note: this week's signal concentrated in US / China / Middle East; no material EU, India, or Japan/Korea news cleared the Scout window.

"Trajectory-based actuator identification via differentiable simulation" (arXiv:2604.10351, updated April 17) demonstrates a calibration method that identifies actuator dynamics from encoder data alone — no torque sensors required — using differentiable physics. Tested on real-robot locomotion, the method improved travel distance by 46% and cut lateral deviation by 75% versus baseline sim-to-real transfer. Applicable to high-gear-ratio actuators, the dominant type in humanoid legs and arms. Directly attacks the actuator-calibration bottleneck every humanoid OEM faces at scale, and lowers bill-of-materials cost by removing the torque sensor.

If you are evaluating humanoid procurement in 2026, two templates now have live production data. AGIBOT shows unit-purchase economics with named-customer scaling commitments and task-level success rates. Agility shows RaaS economics where vendor balance sheet absorbs hardware risk. Both deliver telemetry from real shifts, not demo loops. This is the procurement rubric operators have been waiting for.

Weekly Question: Will AGIBOT hit its Q3 2026 target of 100 G2 units deployed at Longcheer?

Tell us what you think. Reply on X @PhyAIweekly47 or comment on our LinkedIn post.