The Physical AI Dispatch, Issue #7
The Big Picture
Section titled “The Big Picture”This was the week a new class of foundation model arrived for physical AI — and a new class of autonomous warehouse robot. Generalist AI unveiled GEN-1, trained on 500,000 hours of real-world data, achieving 99% task success rates with just one hour of robot-specific adaptation. D-Robotics closed $150M ($270M total) for embodied AI edge computing chips — Chinese silicon for robotics is scaling fast. Locus Robotics launched Locus Array — the first fully autonomous warehouse system combining mobility, picking, and AI vision in one unit, with DHL as early access customer. And Transitive Robotics released its open-source fleet monitoring stack built on ClickHouse, Grafana, and Alertmanager, validating fleet observability as the next infrastructure layer. Models are getting better. Hardware is getting cheaper. The missing piece is production operations.
This Week in Edge AI for Robotics
Section titled “This Week in Edge AI for Robotics”Generalist AI Unveils GEN-1 — 99% Success Rate, One Hour to Adapt
Generalist AI released GEN-1, an embodied foundation model trained on 500,000 hours of real-world data collected from low-cost wearable devices on humans performing millions of activities. The model achieves 99% success rates on tasks where previous models scored 64%, completes tasks 3x faster, and requires only one hour of robot-specific data for adaptation.
The demonstrated capabilities are striking: kitting auto parts for more than an hour, folding T-shirts 86 times in a row, servicing robot vacuums over 200 times, packing blocks more than 1,800 times, folding boxes over 200 times, and packing phones over 100 times in a row. Built from scratch on real-world interaction data rather than fine-tuned from language models, GEN-1 represents a fundamentally different approach to robotic intelligence.
The number that matters most isn’t 99% — it’s one hour. If adapting a robot to a new task takes one hour of data, the rate of model updates across production fleets accelerates dramatically. Every customer deployment can generate custom behaviors quickly, meaning the number of model versions in production grows exponentially. OTA model updates, version tracking, performance monitoring, and rollback capabilities become mandatory at scale. This is the ML ops moment for physical AI.
D-Robotics Closes $270M for Embodied AI Edge Computing
D-Robotics, spun out of autonomous driving chipmaker Horizon Robotics in January 2024, closed a $150M Series B2 round, bringing total Series B funding to $270M. Investors include Didi Global, Prosperity7 Ventures, GL Ventures, Vertex Growth, and 5Y Capital.
The company builds computing platforms and general-purpose infrastructure for consumer robots and embodied AI. Products include the RDK S600 robot computing platform, HoloBrain (perception and decision-making), and HoloMotion (motion planning and control). D-Robotics is building the Chinese alternative to NVIDIA’s Jetson Thor ecosystem — edge compute silicon paired with a full model stack.
Chinese silicon for robotics is scaling fast. Chinese robot OEMs manufacture a significant share of global AMRs and cobots, and many will standardize on D-Robotics hardware. The RDK S600 + HoloBrain + HoloMotion stack is a full-stack alternative that ships with compute, perception, and motion planning out of the box. Hardware-agnostic platforms must support this silicon — or risk being locked out of Chinese-manufactured robot fleets deploying worldwide.
Locus Robotics Launches Locus Array — The First “Robots-to-Goods” System
Locus Robotics launched Locus Array — a fully autonomous fulfillment system combining a mobile robot, integrated picking arm, and AI-powered perception in one unit. It handles picking, putaway, induction, drop-off, and slotting without human intervention, achieving 90% labor reduction.
The system maneuvers within centimeters of double-deep shelving and picks to totes with its gripper system. DHL Supply Chain is an early access customer, with global expansion across Europe and APAC planned. Locus calls this “Robots-to-Goods” (R2G) — a new category distinct from both Goods-to-Person and Person-to-Goods. Named finalist for the MHI Best New Innovation Award at MODEX 2026.
Locus Array is the most operationally complex warehouse robot yet — an AMR, a picking arm, and AI vision in one autonomous unit operating in aisles. The fleet operations requirements for R2G are significantly more complex than simple AMR fleets: manipulation error recovery, grip quality monitoring, path planning in tight aisles, and coordination with human workers. DHL being the early access customer means enterprise-scale deployment is the target from day one. The autonomous warehouse just stopped being theoretical.
Transitive Robotics 2.0 — Open-Source Fleet Monitoring Arrives
Transitive Robotics released version 2.0 of its open-source full-stack robotics framework. The major addition: fleet monitoring and analytics infrastructure built on ClickHouse (time-series data storage), Grafana (visualization), and Alertmanager (custom alerting).
Previously focused on transactional features like WebRTC video streaming, remote teleop, and ROS tools, Transitive is pivoting toward passive monitoring and fleet observability. The company argues that “passive monitoring and alerting need to gradually replace active monitoring by remote operators.” ClickHouse, their chosen database, is already used by OpenAI, Anthropic, and Tesla for high-volume data ingestion.
Fleet observability just got an open-source base layer. Transitive 2.0 applies the cloud-native monitoring stack — metrics, dashboards, alerts — to robot fleets. This validates the thesis that fleet monitoring is the next infrastructure layer the industry needs. What’s still missing in the open-source tier: AI model performance monitoring (not just hardware telemetry), safety compliance, and multi-vendor fleet orchestration. The base layer is open; the enterprise layer is where the business lives.
Robot Density Rising Globally — IFR Reports Growth Across Three Continents
The International Federation of Robotics reported that robot density — units per 10,000 employees — has risen across Asia, Europe, and the Americas. The “World Robotics 2025” data shows factory automation accelerating beyond traditional leaders like South Korea, Japan, and Germany into broader geographies.
The timing aligns with MODEX 2026 (April 13—16), North America’s largest supply chain trade show, which is shaping up as the biggest warehouse robotics showcase yet. Exhibitors include Geekplus (US debut of Geek+ Brain embodied intelligence), Ocado Intelligent Automation (Ocado IQ cloud-based orchestration), Locus Robotics (Locus Array live demos), Quicktron (“one platform for all scenarios”), and SEER Robotics (“all robots, one platform”).
Rising robot density everywhere means fleet management is a global problem, not a regional one. Every incremental robot added to a factory floor increases the complexity of fleet operations — coordination, monitoring, updates, compliance. And the MODEX exhibitor list tells its own story: multiple companies are now pitching “one platform” for warehouse robotics. The platform war for warehouse fleet management is officially underway.
The Numbers
Section titled “The Numbers”Edge Hardware Watch
Section titled “Edge Hardware Watch”Robot computing platform with integrated perception (HoloBrain) and motion planning (HoloMotion) AI models. Edge-cloud architecture for embodied AI. Backed by $270M in Series B funding. Chinese alternative to NVIDIA Jetson Thor ecosystem. Source: Gasgoo
Won 2026 Edge AI and Vision Product of the Year. 99 total platform TOPS across CPU, GPU, and NPU. Production-shipping fifth major edge AI platform for robotics alongside NVIDIA, Qualcomm, AMD, and NXP. Source: Edge AI and Vision Alliance
Edge AI infrastructure deployment across APAC progressing. Edge compute diversifying regionally with new hardware stacks beyond the hyperscaler-dominated data center market. Source: Data Center Dynamics
State space models adapted for compute-in-memory architecture enabling real-time edge inference on video and sensor feeds. University research with longer-term production implications for ultra-efficient robotics perception. Source: Michigan Engineering
Open Source Corner
Section titled “Open Source Corner”Open-source full-stack robotics framework adds fleet observability: ClickHouse for time-series data, Grafana for dashboards, Alertmanager for alerts. Shift from active remote operation to passive monitoring and analytics. Source: The Robot Report
Framework bridging large language models with the Robot Operating System for natural-language robot control. Published in Nature Machine Intelligence. Brings LLM reasoning to real-time robot action. Source: TechXplore
Vision-language-action models for robots, released during National Robotics Week. Enable natural language instruction understanding and physical task execution. Part of the open-source physical AI training stack. Source: NVIDIA Blog
Community call to test the upcoming ROS 2 Lyrical Luth release. Kickoff April 30 at 9am. Production readiness via community validation. Top testers get ROS swag. Source: ROS Discourse
One More Thing
Section titled “One More Thing”There’s a thread connecting every story this week that goes beyond the individual headlines.
GEN-1 can learn new tasks in one hour. D-Robotics is shipping edge chips for a fraction of the cost of Western alternatives. Locus Array operates autonomously in warehouse aisles. Transitive built open-source fleet monitoring.
What ties them together: the cost of deploying a capable robot is falling fast. Cheaper models, cheaper hardware, more autonomous operation, open-source monitoring. Every barrier to scale is being removed — except one.
The operations layer — the software that keeps thousands of these robots running correctly in production, across vendors, across sites, across model versions — is still being built. And the window to build it is narrowing, because at the rate GEN-1 adapts and factories manufacture, the fleet is growing faster than anyone expected.