Isaac ROS 5.0 lands on Pixi

Today: Isaac ROS 5.0 becomes installable with Pixi, a 617 hour motion capture dataset ships a humanoid benchmark, paid home repairs turn into robot training data, ros2_control 6.12.0 removes APIs deprecated since Jazzy, and retail shelf robots move from pilots to fleets.

Isaac ROS 5.0 moves into the Pixi channel

NVIDIA shipped Isaac ROS 5.0 on 22 September with a version tag on the isaac_ros_common repository. The packaging is the new part: prefix-dev, the team behind the Pixi package manager, published isaac-forge, which rebuilds Isaac ROS as conda packages on top of RoboStack so the same GPU accelerated stack installs on x86_64 Linux and on Jetson ARM64 (ROS Discourse).

The version split matters. Isaac ROS 5.0 needs the ROS 2 buffer API that arrives with Lyrical, because its GPU image and tensor pipeline exchanges rosidl Buffer message fields that Jazzy messages do not have. NVIDIA publishes 5.0 for Lyrical only, while the maintenance channel isaac-forge/jazzy freezes at 4.6 with Python 3.12 and the older flat channel is frozen, so pin 4.6.* on Jazzy and 5.0.* on Lyrical.

Coverage is partial: cuVSLAM, nvblox, the cuMotion controllers and teleop do not build for Lyrical yet. On Jetson the packages target JetPack 7, glibc 2.38 and CUDA 13 user space libraries, with Orin Nano selected through SM87, and Pixi installs the CUDA libraries but not the host driver.

For developers the appeal is reproducibility: one pixi run replaces the apt and Docker layering that made Isaac ROS environments hard to reproduce, and the repository ships examples for TensorRT YOLOv8 inference on an image, video, webcam or RTSP stream plus Grounding DINO open vocabulary detection. The Jetson Orin Nano and Orin NX comparison covers the compute side and the Isaac Sim primer the simulation side. Watch whether NVIDIA adopts the channel officially.

A 617 hour motion dataset for humanoid policies

HiPHI, by Jiahao Ji and 14 co-authors, is a 600+ hour whole-body human motion dataset built with optical motion capture and accepted at CoRL 2026. The arXiv abstract reports sub-millimeter marker tracking for full body motion and mesh level object trajectories, with collection guided by FrameNet, a linguistic framework for organizing human actions, so coverage is chosen systematically.

A companion white paper puts numbers on the release: 617.5 hours of whole-body motion, of which 245.7 hours include human-object interaction with synchronized object trajectories and meshes, plus reinforcement learning policies trained on the data and deployed on a physical Unitree G1 humanoid. Treat those as vendor numbers: the white paper is sponsored by Noitom Robotics, which also hosts the project page.

The benchmark suite is the part developers can use before trusting a policy: it scores motion space diversity, interaction grounding, object consistency and physical AI applications. It addresses the data bottleneck: internet video lacks precise physical state, lab mocap is accurate but narrow, and policies trained on one embodiment meet the world model generalization problem when the task changes. Watch for independent reproduction of the reported sim-to-real transfer.

Home repairs become robot training data

HUMXN, an AI data company, launched a program on 6 October that pays for qualifying plumbing, electrical, HVAC, air duct cleaning and carpet cleaning appointments in Minneapolis, Chicago and Miami, with technicians wearing capture hardware during normal service calls (HUMXN). Homeowners pay nothing, service companies are paid for the work, and HUMXN converts the recordings into datasets for AI and robotics groups, which founder Thomas Unise frames as an underused source of demonstrations of perception, tool use and decision making.

The claim is plausible and the evidence is thin: HUMXN has published no dataset, named no robotics customer and shared no capture specification, so this is an announcement rather than a proven data supply. The shape of the problem is what makes it worth tracking. A plumber diagnosing a leak produces exactly the long horizon, tool centred behaviour that lab teleoperation struggles to scale, the same gap described in the two halves of the data bottleneck. Watch for a named buyer or a published sample.

ros2_control 6.12.0 removes deprecated APIs

The ros-controls maintainers warned that ros2_control 6.12.0 removes the API deprecated since Jazzy, has already been released into Lyrical and Rolling, and will reach users with the next sync (ROS Discourse). Raw pointer handles are gone, so a StateInterface or CommandInterface can no longer be constructed from a double* pointing into your own member variables: handles now own their data and are reached through shared pointers and the getter and setter API. The legacy export_state_interfaces() and export_command_interfaces() exports that returned interfaces by value are removed, as are the by-value export methods on ChainableControllerInterface, replaced by shared pointer on_export_*_interfaces_list() callbacks (ros2_control#3610).

Jazzy and Kilted keep the old API. If your hardware component or controller builds with deprecation warnings from hardware_interface or controller_interface, assume it fails after the sync and read the migration notes first. As with executors and callback groups, the interfaces behind a controller change under you and warnings are the only advance notice.

A recycling robot bets on a pincer and a payback number

Danu Robotics, a six year old Edinburgh company founded by Amy Ma, is bringing its H.E.R.O. recycling sorter to market with $5 million in late seed funding. The robot uses a pincer claw rather than the suction systems used by competitors such as Glacier and RecycleEye, and the company says it continuously retrains the software from the sorting line (TechCrunch).

The commercial claims are specific and unverified: $500,000 of signed contracts, letters of interest from two large customers, more than 200 prospects in the pipeline, and a payback story of $485,000 in added annual revenue per site against a $160,000 installation and $24,000 a year in maintenance, in a market Ma puts at $20 billion across Europe and North America. The near term test is whether the signed contracts become running lines, the same pilot to fleet gap that showed up in warehouse deployments.

Retail shelf robots move from pilots to fleets

At Groceryshop 2026, Simbe Robotics reported it has completed a rollout of 3,000 Tally shelf scanning robots across retailers and countries, while Badger Technologies described itself as evolving from a robotics company into a retail intelligence company under new chief executive John Gehre, with Jabil manufacturing behind it (The Robot Report).

The split is deployment economics. ShelfOptix targets drug, convenience and dollar stores with a foldable robot escorted between locations by a human instead of one robot per store, which spreads the cost of recurring audits over a wider footprint. Blue Collar Robotics, a startup competition finalist, goes after the harder case of moving products between shelves and carts, which needs navigation, vision, picking, controls and back end integration to work together. MUSE keeps adding modules to one mobile platform, from customer guidance to replenishment and cleaning. The safety layer for AMRs sharing store space is tightening too, as the user side AMR safety standard shows.

Quick Bites

• Draft REP 158 proposes OpenUSD conventions for simulation asset interoperability, so one asset works across Gazebo, O3DE, Isaac Sim, Newton, MuJoCo and Genesis. It mandates MKS units and Z-up coordinates, adds declarative ROS interface schemas (RosContextAPI, Topics, Services, Actions, Frames) and requires down conversion to glTF 2.0 and URDF (ROS Discourse).
• dynrec 0.2.0 lets a rosbag2 recording change its topic set at runtime without stopping the writer, and adds disk and size guards, simulation time support and per topic loss reporting. The format stays compatible with ros2 bag play and ros2 bag info, and the author is still seeking feedback before proposing it upstream (ROS Discourse).
• Jumper is an open source crab robot, hardware and software, written in Rust with a browser based simulator, and it does not run ROS yet (GitHub).
• The FCC added foreign produced advanced robotic devices to its covered list in July, blocking new devices from authorization to be imported, marketed or sold in the U.S. The Robot Report argues the change pushes inference toward the edge and strengthens the case for small language models fine tuned on facility data (The Robot Report).