CircuitHub's Grid, A3 orders, world model tests

This week: a PCB factory built as one machine, the strongest order quarter of the year, and four developer reports from real hardware and simulation.

CircuitHub turns PCB assembly into one robot

CircuitHub published details of The Grid, a factory-scale robot in Massachusetts that assembles circuit boards on demand (CircuitHub). Levitating movers carry each board and its parts between stations that print solder paste, place components, reflow, solder and inspect, while learned models, AI agents and classical optimizers handle scheduling. Every board takes its own route, so no changeover exists between designs.

The company says the system has run for customers for five years and produces tens of millions of dollars of electronics a year. Across its service it counts 2.2 million boards shipped and 150 million parts placed, and it plans to 5x capacity this year.

Its own comparison prices a 4-layer motor controller, 118 parts per board (38 unique), five boards delivered in the US: $630 on The Grid in 4 days against about $465 and a 9 day median from a Shenzhen quick-turn shop once shipping and tariffs are added. Vendor quote, not an audited benchmark, but the right comparison for hardware teams.

Robot builders care because motor controllers and sensor boards are low volume and high mix, and this is the step that sets iteration speed.

A3 reports the strongest order quarter of 2026

North American companies ordered 10,989 robots worth $723.5 million in the third quarter, up 24.8 percent in units and 26.1 percent in value year over year, according to the Association for Advancing Automation (The Robot Report). Nine month orders reached 28,984 units and about $1.89 billion, up 9.6 and 13.3 percent.

The mix moved more than the total. Automotive OEM orders returned to growth at 7.5 percent after a weak first half, though they remain 17 percent below the year earlier level through September. Automotive component suppliers rose 90.4 percent, semiconductors and electronics 97.6 percent, plastics and rubber 67.7 percent, life sciences 38.7 percent. Food and consumer goods fell 23.7 percent. Collaborative robots accounted for 1,831 units and $63.8 million, up about 56 percent.

A3 executive vice president Alex Shikany attributed the quarter to breadth rather than one sector. The electronics figure matters for developers: the packaging and inspection demand behind those orders is what absorbs NVIDIA modules, vision sensors and embedded boards.

WorldOracle scores world models against real executions

A benchmark from Princeton, UT Austin, Toronto, NVIDIA and AgiBot gives each action-conditioned world model the same initial observation and low-level action sequence as a real episode, then scores the predicted rollout against the actual execution (arXiv 2610.18826, The Robot Report).

The result to internalize is that arm motion is easy and object state is not. On AgiBot G2 human-operated data, IRASim scored 78.8 for the robot arm but 29.1 for the manipulated object, and Cosmos 3 scored 89.4 and 45.4. Over full sequences rather than the first 30 seconds, scores collapsed: DreamDojo fell from 50.56 to 27.08 and Cosmos 3 from 53.99 to 26.60, while GE-Sim 2.0 held best, from 73.86 to 63.33.

Four metrics feed the single view score: physical state correctness, interaction spatial alignment, key interaction area fidelity and, for multi-view models, action coherence. Failed real executions are test cases, and predicting a failure correctly counts as correct, which makes confident hallucinated success measurable.

The platform opened to developers on October 16. If a policy is trained or planned against a world model, arm accuracy alone is not evidence: the conditioning choices need checking at the object.

slam_toolbox multi-robot SLAM on three TurtleBots

A developer ran slam_toolbox’s decentralized multi-robot node with one Nav2 stack per robot on three TurtleBot3 Burgers under ROS 2 Jazzy (ROS Discourse, code). The three robots build one shared map, and each sees the other two as costmap obstacles.

The friction list is the value. The multi-robot node is not in the Jazzy release, so the commit from PR #592 had to be cherry-picked and built from source. The LDS-02 lidar returns a different beam count per scan and slam_toolbox rejects variable length scans, so every scan is resampled to 360 beams. The node expects a per-robot TF tree on /robotN/tf with plain frame names plus a shared global_odom, so namespaced transforms must be relayed into that layout. On Jazzy, navigation_launch.py does not push the namespace onto Nav2 nodes, and turtlebot3_node needs enable_stamped_cmd_vel to accept TwistStamped.

The network finding is the practical one: with three Nav2 stacks over WiFi, odometry dropped from 20 Hz to 1 to 4 Hz until Fast DDS was bound to the laptop’s LAN address and socket buffers were enlarged. Start poses must also land within about 0.25 m and 20 degrees of the real layout. Single robot Nav2 setups hide all of it.

Zenoh plus RaptorQ for links that drop a quarter of their packets

An open source prototype pairs Eclipse Zenoh with RaptorQ fountain codes for outdoor ground robots that lose 15 to 30 percent of packets under foliage, structures and terrain (ROS Discourse, code).

It targets head-of-line blocking, where a retransmitting map or boundary file pushes urgent commands and telemetry behind it. Bulk assets are compressed, split across interleaved RaptorQ source blocks (RFC 6330) and reconstructed once enough droplets arrive, with no round trips. Map streaming is paced by a token bucket at a default 128 kbps using Zenoh’s low priority and drop congestion control, while signed Ed25519 commands take high priority with blocking delivery.

The published demo runs two peers over a simulated 25 percent loss link: 158,500 bytes compressed to 36,042, split into 3 source blocks and 111 droplets, of which 83 arrived and 28 were dropped. All 3 blocks solved in 3.82 s with zero retransmissions. That is simulation, not field data, and the author is asking whether this belongs in a transport daemon or a native ROS 2 node. Underneath sits the QoS question of which traffic may starve which.

MS-Human-700 puts a 700 muscle body in MuJoCo

A research group published MS-Human-700, a whole-body human musculoskeletal model for MuJoCo with 90 segments, 206 joints and 700 muscle-tendon units, plus controllers trained on it (project page). Training uses MuJoCo Warp for parallel GPU simulation.

The model is the reference artifact: anatomically plausible parameters, published with work spanning ICRA 2024 for the model and hierarchical reinforcement learning, ICLR 2025 for hierarchical model predictive control, ICLR 2026 for value-guided exploration, and IROS 2026 for using the body as a digital human surrogate to co-optimize an exoskeleton’s structure and controller.

For robotics the use is evaluation. Joint targets are a fiction next to 700 muscles, and an exoskeleton that looks fine against a rigid mannequin can be wrong against a body that generates force physiologically. The same overactuation problem appears on the robot side in action space design.

Quick Bites

• A Bolt delivery robot entered a Tallinn crossing on a red light and collided with a bus; up to nine standing passengers were injured when the bus braked, and the robot left before police arrived (ERR).
• Audi is testing Agile Robots dual-arm humanoids on wheeled bases for parts sequencing at Ingolstadt, with NVIDIA software and a Dassault Delmia virtual twin used to check feasibility, timings and ergonomics before deployment (Assembly).
• Warp Sciences left stealth with $215 million led by Lightspeed to build a US chemistry contract research organization combining chemists, robotics and AI, starting with lead optimization analog synthesis (The Robot Report).
• A ros2_control issue asks whether command and state interfaces should carry strings for firmware versions, interlock checksums and serial numbers, against the real-time constraints of the hardware loop (GitHub).
• RoboStack moved all ROS distributions into one monorepo with a bot that opens package addition pull requests (ROS Discourse).