The Physical AI Dispatch, Issue #6
The Big Picture
Section titled “The Big Picture”This was the week physical AI graduated from prototype to production. Japan declared a national target of capturing 30% of the global physical AI market by 2040, backing it with industrial policy and SoftBank-led deployments. Geekplus reported its first-ever profit — 31.6% revenue growth with 75% international revenue — proving that warehouse AMR fleet economics work at scale. Figure AI ramped BotQ to one humanoid per hour, a 24x manufacturing leap, with 350+ Figure 03 units shipped and a 12,000-unit annual target. Sanctuary AI achieved zero-shot sim-to-real transfer with a 21-DOF hydraulic hand, reorienting a cube 10 times without dropping it. And National Robotics Week kicked off with NVIDIA releasing Newton 1.0, Isaac GR00T open models, Isaac Sim 6.0, and Isaac Lab 3.0 — making the physical AI simulation and training stack open-source and production-ready. The robots work. The factories are scaling. The simulation stack is open. The question now is: who runs them in production?
This Week in Edge AI for Robotics
Section titled “This Week in Edge AI for Robotics”Japan Goes All-In on Physical AI — METI Targets 30% of the Global Market by 2040
Japan’s Ministry of Economy, Trade and Industry formally set a target of capturing 30% of the global physical AI market by 2040. It’s a national strategy built on Japan’s existing dominance: Japanese manufacturers account for approximately 70% of global industrial robot production. But the bet isn’t just on hardware — it’s on layering AI onto that installed base.
SoftBank is already combining vision-language models with real-time control systems to enable robots to interpret environments and execute complex tasks autonomously. Investment is shifting toward orchestration software, digital twins, simulation tools, and integration platforms. Japan’s labor shortage — structural and intensifying — makes this more than industrial policy. It’s a demographic necessity.
Japan has the installed base, the demographic urgency, and now the policy mandate. When a government puts a number on it — 30% — funding, procurement, and regulatory frameworks follow. The orchestration software layer that connects AI to Japan’s massive existing robot infrastructure is the opportunity. This isn’t about building new robots; it’s about making the ones that already exist smarter and more coordinated.
Geekplus Hits Profitability — First Profitable Major AMR Company
Geekplus (HKEX: 2590.HK) reported FY2025 results on March 31: revenue up 31.6% year-over-year to 3.171 billion RMB, gross profit up 34.4% with margins improving to 35.5%, and its first-ever adjusted net profit of 43.8 million RMB. Total orders surged 31.7% to 4.137 billion RMB. International revenue accounts for 75.3% of the total at 46.6% gross margins.
The engine behind the numbers is Geek+ Brain — the company’s proprietary embodied intelligence platform that powers adaptive robotic picking across ultra-large-scale warehouse environments. The platform is central to Geekplus’s expansion in the Americas, where new signed orders grew more than 50% year-over-year in 2025.
Profitability in warehouse robotics isn’t just a business milestone — it’s a market validation. Geekplus proving that AMR fleet deployments generate sustainable returns at scale means procurement teams can now justify fleet investments with real financial evidence, not projections. The Geek+ Brain platform is evolving from picking optimization to a broader embodied intelligence layer. What’s missing on top: operational intelligence across the fleet — monitoring, diagnostics, and compliance at scale.
Figure AI Reaches One Robot Per Hour at BotQ — 24x Manufacturing Leap
Figure AI’s BotQ manufacturing facility in California increased production from one robot per day to one per hour — a 24x throughput jump achieved in under four months. Over 150 networked workstations and custom production software enabled the delivery of more than 350 Figure 03 humanoid units.
The quality numbers are equally notable: end-of-line first-pass yield exceeds 80%, battery production at 99.3% yield, and over 9,000 actuators produced. Each robot undergoes 80+ functional tests, including stress and burn-in exercises like squats and jogging. Annual capacity target: 12,000 units, with a four-year goal of 100,000 robots per year.
Manufacturing humanoid robots at scale was supposed to be the hard part. Figure just compressed the timeline dramatically. But production and deployment are different problems. At 12,000 units per year shipping to enterprise customers across multiple industries, the fleet operations challenge — firmware updates, remote diagnostics, safety compliance, performance monitoring — becomes the bottleneck. Manufacturing solved. Operations is next.
Sanctuary AI Achieves Zero-Shot Sim-to-Real Dexterity
Sanctuary AI demonstrated a milestone in sim-to-real robotics: its 21-degree-of-freedom hydraulic robotic hand autonomously reoriented a lettered cube to match a specified goal — 10 consecutive times without dropping it. The manipulation occurs entirely at the fingertips without palm support, requiring the hand to simultaneously grasp and reorient the object.
The key technical achievement is zero-shot transfer. The manipulation policy was trained entirely in simulation and deployed directly to the physical hardware with no real-world fine-tuning. Sanctuary AI’s hands have 21 degrees of freedom — enabling finger abduction and sophisticated in-hand manipulation that most robotic hands cannot perform.
Zero-shot sim-to-real transfer changes the deployment model for robot behaviors. If policies can be developed entirely in simulation and deployed directly to production hardware, the iteration cycle from training to fleet update accelerates dramatically. But the faster you can create new behaviors in simulation, the more critical the deployment pipeline becomes — how do you push updates to production fleets safely, monitor performance in the field, and roll back if something goes wrong? The sim-to-real pipeline needs operational guardrails.
National Robotics Week: NVIDIA Ships the Open-Source Physical AI Stack
National Robotics Week 2026 (April 4—12) launched with NVIDIA delivering a comprehensive wave of physical AI tooling.
Isaac GR00T open models enable robots to understand natural language instructions and perform complex, multistep tasks using vision-language-action reasoning.
Cosmos world models generate synthetic data and training scenarios at scale, helping robot systems learn more efficiently and generalize across environments.
Newton 1.0 (developed with Google DeepMind and Disney Research, managed by the Linux Foundation) reached general availability — a GPU-accelerated open-source physics engine clocking 475x faster than MJX on RTX PRO 6000 Blackwell GPUs.
Isaac Sim 6.0 and Isaac Lab 3.0 also hit GA, providing production-ready simulation and reinforcement learning frameworks built on NVIDIA Omniverse.
NVIDIA just made the complete training and simulation stack for physical AI open-source and production-ready. Any team can now build, train, and simulate robot behaviors at scale. But the gap between “trained in simulation” and “running in production across a fleet” is where the real infrastructure challenge lives. NVIDIA builds the development stack. The deployment and operations stack is what comes next.
The Numbers
Section titled “The Numbers”Edge Hardware Watch
Section titled “Edge Hardware Watch”Winner of the 2026 Edge AI and Vision Product of the Year. Up to 16 cores, Intel Arc GPU with 8 Xe cores, upgraded NPU reaching 99 total platform TOPS — nearly 3x the Series 1. Fifth major silicon platform hitting 80—100+ TOPS for edge robotics. Source: Edge AI and Vision Alliance
University of Michigan developed a hardware-software co-design adjusting state space models for compute-in-memory architecture. Enables real-time video and sensor processing on edge devices with high energy efficiency and low latency. Promising for next-generation robotics perception. Source: Michigan Engineering
University of Cambridge researchers created a modified hafnium oxide memristor replicating neuron communication patterns. Could reduce AI energy consumption by 70%. Long-term play for ultra-efficient edge AI in resource-constrained robot deployments. Source: ScienceDaily
US-based AI chip startup Blaize and Chinese hardware developer NeoTensr signed a contract worth up to $50M for edge AI infrastructure deployment across APAC. Edge compute going regional at scale with new hardware stacks. Source: Data Center Dynamics
Open Source Corner
Section titled “Open Source Corner”GPU-accelerated open-source physics engine from NVIDIA, Google DeepMind, and Disney Research. 475x faster than MJX on Blackwell GPUs. Ships with MuJoCo Warp and Kamino solvers, deformable simulation, SDF collisions, and hydroelastic contact. Integrates with Isaac Lab 3.0. Source: GitHub
New framework integrating large language models with the Robot Operating System for natural-language robot control. Published in Nature Machine Intelligence. Bridges the gap between LLM reasoning and real-time robot action. Source: Open Source For You
Community call to test the upcoming ROS 2 Lyrical Luth release. Kickoff April 30 at 9am. Production readiness depends on community validation — top testers get ROS swag. Source: ROS Discourse
Presentation proposals for ROSCon Global 2026 (Toronto, Sept 22—24) due April 26. Physical AI infrastructure — fleet monitoring, sim-to-real deployment, edge diagnostics — is a hot topic across the 2026 ROSCon circuit. Source: ROSCon 2026
One More Thing
Section titled “One More Thing”There’s a pattern forming that’s easy to miss if you focus on any one story in isolation.
Japan set a national target for physical AI market share. Geekplus proved that robot fleet economics generate profit. Figure showed that humanoid manufacturing can scale. Sanctuary AI demonstrated that robot behaviors can transfer directly from simulation to hardware. And NVIDIA open-sourced the simulation and training stack.
Every layer of the physical AI stack is reaching production maturity — except one. The development tools are open. The hardware is shipping. The manufacturing lines are running. The business models work. What’s still being built is the operations layer: the software that deploys trained models to production fleets, monitors them in the field, pushes updates safely, and maintains compliance across thousands of units.
That layer is where the next great infrastructure company will be built. And the clock is ticking — because at one robot per hour, the fleet is growing fast.