ISO 21423 gives mixed fleets a language

ISO 21423 is published, and an open source fleet platform already claims a reference implementation. Also this week: a four winch cable robot, Bazel builds for ROS 2 Lyrical, a learned inverse kinematics layer, and NVIDIA’s city scale physical AI pitch.

ISO 21423 covers mixed mobile robot fleets

The Association for Advancing Automation has published ISO 21423:2026, an international standard for communications among industrial mobile robots from different vendors (The Robot Report). It was developed in ISO Technical Committee 299 and covers automated guided vehicles, autonomous mobile robots and the fleet managers in front of them.

The payload of the standard is operational state: location, speed, direction, condition, tasking and availability. It also reaches past the fleet to machines tended by robots, ERP systems, building systems such as lifts and automated doors, and people carrying mobile devices. Carole Franklin, A3’s director of standards development, said companies want to choose the robots that fit their applications and have those robots work together on the same floor. A3 calls it the first non-safety robot standard it has helped produce.

InOrbit released OpenRobOps on 15 September, an Apache-2.0 fleet operations platform the company describes as the first ISO 21423 reference implementation (repository, A3). The repository covers ROS 2, Open-RMF and the standard, and it was pushed again on 3 November.

VDA 5050 already specifies a fleet interface for AGVs and AMRs, with its official specification on GitHub. ISO 21423 is wider, describing the building and the people as well as the vehicles.

For developers the work lands in the fleet adapter. Deadlock, zone reservation and task handover are the multi-robot version of the path planning problem Nav2 solves for one robot, and a standard only helps if both sides populate the same fields with the same meaning. Watch for a conformance suite.

A cable robot that uses the ceiling instead of an arm

Neufangled Robotics built Stringman, a cable driven parallel robot for one room: four winch driven lines anchored near the ceiling corners hold a 638 g gripper, positioned by paying line in and out (Neufangled).

Work area is a room 3 to 10 m on the XY diagonal with 2.3 to 4 m ceilings. Payload is up to 2.2 kg near the floor and 1.9 kg on average at a 28 N line tension limit, net of the gripper, at about 0.4 m/s. Compute is external: a PC runs the controller, with a Pi Zero 2W in the gripper and Pi 3A+ boards at the anchors, over 5 GHz Wi-Fi.

The project holds OSHWA certification US002788, everything is Apache-2.0, and the PCBs and firmware are on GitHub (PCBs, firmware). There is a Python API and an MCP server, and the vendor says models run on the owner’s hardware with no data leaving the network.

The teleoperation data is public: a LeRobot dataset with 16 episodes at 30 fps from four 512x512 cameras, an 11 dimension action vector and a 44 dimension state that includes all four cable tensions, gripper pose, a laser rangefinder and finger pressure. Tension is the control problem, not six joint angles.

Neufangled’s table puts average payload at 1.9 kg against 0.32 kg for a 0.75 m serial arm on the same actuators, and work volume per dollar about 500 times higher. Company figures, and the arm number ignores link weight. Repeatability and settling time under load are not published, and 16 episodes is small. See the robot data bottleneck.

Bazel builds arrive for ROS 2 Lyrical

The ROS Central Registry shipped a new Lyrical release and a call for a working group (ROS Discourse). The registry is a community Bazel build of ROS from Lyrical onward, and the post is explicit that it complements colcon and pixi rather than replacing them.

The index at bazel.ros.org lists 1,991 modules across 6,889 versions from 53 maintainers, with rosdistro at lyrical.2026-10-05.rcr.1. Coverage includes message pipelines, nodes, launch, rviz2, MoveIt, Nav2 and Gazebo.

The build model is the point. Packages are Bazel modules for Bzlmod federation, with hermetic C++, Python and Rust toolchains, so a build does not depend on what is installed on the machine. Intrinsic provides a read-only GCS cache that CI warms on every commit to main, and a .bazelrc edit turns it off. It builds on the Bazel team’s work and Milan Vukov’s rules_ros2, particularly IDL handling for message dependency chains.

It is still a proof of concept; Andrew Symington, who relayed the release, plans a working group. Contributors get AGENTS.md and a .agent/skills folder in the registry repository, and CI rejects breaking changes. A remote cache is welcome after a cold colcon build, though a second build system is also a second maintenance surface. Background: what ROS 2 actually is.

A learned inverse kinematics layer for arms

Analytic inverse kinematics is exact in theory and fussy in practice: redundancy, singularities and the offset a position controller needs at rest show up as millimetre errors. PolicyIK replaces the solver with a goal conditioned reinforcement learning policy that updates joint position references from pose feedback (arXiv 2611.05372, submitted 3 November, from NYU and NYU Abu Dhabi).

The policy observes its accumulated joint reference relative to measured joints, so it can hold the offset the low level controller needs at rest. Pose error is split into two scales with an error dependent action gain, fast for large motions and fine at the last millimetre, and progress plus dwell rewards teach it to hold the tolerance.

Reported numbers: a single PPO attempt holds mean error over the final second within 1 mm and 0.2 degrees on 93.5 percent of held out tasks, and a deployment protocol lifts that to 98.9 percent. Trained on a physical arm with an off-policy learner it reports higher reaching and holding success than four classical IK controllers at equal time budgets, and in simulated pick and place it executes end-effector commands from the pi_0.5 VLA with 95.7 percent success.

These are author results without independent replication, and real-arm training costs robot time. The architectural point holds: a VLA outputs a pose, and the layer between that pose and a joint command decides precision, which is why VLA models alone are not enough. Code at Tinaliu0123/PolicyIK.

NVIDIA pushes physical AI into city infrastructure

NVIDIA put its city deployments in one post on 3 November, framed as physical and agentic AI rather than robots (blog). The numbers are vendor reported.

In Malaysia, ITMAX runs SamurAI V2 vision agents on RTX PRO 6000 GPUs, targeting more than 17,000 cameras by 2027, and uses TeraSim, an open source traffic simulator from Laplace Intelligence, to test signal timing, which NVIDIA credits with cutting peak hour corridor travel time by up to 26 percent. In Wrocław, K2K builds shared city infrastructure on Cosmos world reasoning models and the Metropolis video search blueprint.

The reusable layer is the plumbing: video ingest, multi camera tracking and a world model on top, on the same GPU stack that runs simulation. A city camera network is a large scale test of the perception pipeline a mobile robot fleet needs indoors. Whether the travel figures survive an independent audit is open.

Quick Bites

• Christophe Bédard is the ROS Boss for ROS 2 Makoa Mata-mata, planned for May 2027, a pick made more than six months ahead after the Lyrical retrospective recommended choosing earlier (ROS Discourse).
• WiBotic CEO Ben Waters argues that as fleets grow, opportunity charging in idle moments matters more than parking robots at wall chargers, and charging data can surface power problems before they cost uptime (The Robot Report).
• Trucks VC fielded pitches from 81 autonomous vehicle startups in 2026, two thirds more than 2025, and partner Puneeth Meruva says foundation models let some teams deploy within months on less than $1 million raised (TechCrunch).