This was the week the robot stopped being the product. Policies, perception primitives and procurement curves moved independently of the metal they ride on, and each one crossed a commercial threshold the same seven days. The platform-vs-model debate that dominated 2025 collapsed into a single question -- which brain, bolted to which body, sold at which sticker -- and the answers now span consumer driveways, industrial inspection bays and public roads. Robotics just became a competitive market with real SKUs, and the hardware is no longer the interesting variable.
Watch & Listen First
London, meet Waymo -- autonomous miles begin on UK roads | Waymo Blog
April 14 post on Waymo's first autonomous runs across London's 100-square-mile pilot. Trained specialist still behind the wheel, but the I-Pace fleet is driving itself.
Embodied AI 101 -- Daily briefings on physical intelligence | Apple Podcasts
This week digs into compositional generalization and Gemini Robotics-ER 1.6.
Key Takeaways
- Stop buying task-specialist policies on one-year contracts. Generalist VLAs are now matching specialists on the same manipulation benchmarks, and the fleet of narrow models every integrator quietly maintains is about to become a liability. Renegotiate around policy-swap rights, not task lock-in.
- Benchmark BOM, not demo reels. Automotive tier-1 suppliers are now building humanoids at car-adjacent sticker prices, which resets every procurement spreadsheet written against US startup quotes. If your 2026 capex assumes $150K per unit, cut it by two-thirds and re-run the payback math.
- Industrial inspection is the first lights-out beachhead. Gauge and sight-glass reading has blocked unmanned plant operation for twenty years; the perception primitive now ships in a general API. Audit which of your inspection rounds can move to a quadruped this quarter.
- Research procurement just collapsed to a GPU-rig line item. Full-size humanoids with LiDAR are listing at workstation pricing, and whole-body controllers are pulling locomotion toward biological speeds through software alone. University and corporate labs without a humanoid on the floor by end of Q2 will be shopping into a shortage.
- Data efficiency, not data scale, is the new moat. One-hour-per-task finetuning on wearable-captured human activity is beating multi-month teleop campaigns. Your robot-data strategy needs a wearables arm before it needs another cell.
The Big Story
Physical Intelligence unveils pi0.7, a steerable foundation model that generalizes to unseen tasks · April 16, 2026 · TechCrunch
-> pi0.7's core claim is compositional generalization -- the model recombines skills learned in different contexts to solve tasks it never saw in training, rather than rote-matching demos. On coffee, laundry, and box assembly it matches PI's own specialist VLA models, and co-founder Sergey Levine says capability is scaling "more than linearly" past a data threshold. First credible argument that one generalist policy can replace the fleet of narrow VLA specialists every humanoid company has been quietly maintaining. The startup is reportedly raising at an $11B valuation, up from $5.6B.
Also This Week
Chery begins selling a $42,000 humanoid to Chinese consumers · April 13, 2026 · Robotics & Automation News
-> First humanoid on an automaker's consumer order sheet at car-adjacent pricing, sold through Chery's Aimoga brand on JD.com alongside a $15,800 quadruped. Chery is leveraging the same tier-1 suppliers, casting lines and motor plants that build its EVs -- which is how the BOM lands at $42K while US startups still quote $150K+. The humanoid pricing curve just started to look like the EV pricing curve.
DeepMind ships Gemini Robotics-ER 1.6 with instrument-reading built in · April 14, 2026 · DeepMind Blog
-> New embodied-reasoning model exposed via the Gemini API and AI Studio, with sharper spatial reasoning, success detection, and "instrument reading" that hits 93% on gauges and sight glasses. DeepMind built the inspection capability in collaboration with Boston Dynamics for Spot deployments. One VLA, many bodies, now with the specific perception primitives industrial inspection has been waiting on.
Generalist AI releases GEN-1 foundation model with 99% task success · April 11, 2026 · Robotics & Automation News
-> Trained on large-scale human-activity data captured via wearables, GEN-1 reports 99% success where prior models hit 64%, a 3x speedup, and ~1 hour of new robot data per task. Early access only, not a shipped product -- but a serious claim that data efficiency is cracking, not just model size.
Unitree G1 on sale at $16,000; H1 hits 10 m/s on the track · April 9-11, 2026 · RoboHorizon · Humanoids Daily
-> A 23-DoF humanoid with 3D LiDAR at GPU-rig pricing resets the research procurement curve. The H1's 10.1 m/s sprint -- up from 3.3 m/s in 2024 -- approaches Usain Bolt's average pace and, per Unitree, comes from software and control changes rather than new hardware.
Waymo begins autonomous driving in London · April 14, 2026 · TechCrunch
-> ~100 I-Pace units across roughly 100 square miles, Waymo's first European pilot, with trained specialists still behind the wheel pending UK regulatory sign-off. Competing head-to-head with Wayve and Uber for London's first commercial robotaxi launch.
From the Lab
Squint: Fast Visual RL for Sim-to-Real Robotics · Christensen Lab, UCSD
-> Visual Soft Actor Critic combining parallel sim, a distributional critic, resolution "squinting" and a tuned update-to-data ratio. Trains manipulation policies in ~15 minutes on a single GPU across 8 ManiSkill3 tasks, with transfer to a real robot. Per-deployment finetuning from pixels is finally plausible.
Learning Versatile Humanoid Manipulation with Touch Dreaming (HTD) · arXiv 2604.13015
-> A multimodal Transformer that predicts future tactile latents alongside action chunks, paired with an RL whole-body controller and VR teleop. 90.9% relative success-rate gain over the strongest baseline across five contact-rich tasks. For contact-rich manipulation, the claim that touch is the core signal is starting to look correct.
Worth Reading
- Cadence and NVIDIA expand physics-simulation partnership for sim-to-real -- April 16. High-fidelity multiphysics folded into Isaac -- a direct attack on the sim-to-real gap in contact dynamics and compliant materials.
The policy now matters more than the platform. A two-year-old startup's model is reshaping how generalist robots get trained, a DeepMind model landed inside Spot, and a carmaker put a humanoid in a driveway for $42K. The platform wars have flipped into model wars -- and the robots are the distribution channel.