Helix 2.5, OpenArm, Arm capability levels

A humanoid policy generalising to homes it has never entered, an open humanoid arm with published CAD, a vendor coalition defining what robot capability means, and embedded ROS 2 without a translation agent.

Figure’s Helix 2.5 works in homes it has never seen

Figure published Helix 2.5 on 17 September, its most advanced neural network to date. It is pretrained on Index, Figure’s dataset of human behaviour, and from that one base model Figure derived three whole body behaviours: tidying living rooms, folding towels, and making beds. The robot then went into 30 Bay Area homes with zero data collected in any of them, and no fine tuning or adaptation to the rooms or the objects.

The numbers matter more than the demo. Holding task data, architecture, training and evaluation fixed, Index pretraining alone moved zero shot success from 9% to 56%. A behaviour built this way used half the task specific data of a representative Helix 02 behaviour while generalising across 30 unseen homes instead of one. Figure also reports that repeatedly doubling the Index pretraining data improved robot action prediction smoothly enough to forecast the largest run’s loss to four decimal places before training.

All of it is vendor reported, with no independent replication, and Figure writes that general humanoid robotics is not solved. The shift is in the data argument: if human experience transfers to whole body behaviour, the bottleneck moves from teleoperation hours per environment to scaling human data and the fine tuning step. Background: why VLA models matter.

Source: Figure.

OpenArm ships humanoid arm hardware under an open licence

OpenArm reached Hacker News on 13 September with 214 points, and the repository now sits at 3.4k stars and 373 forks across 452 commits. It is a 7 degree of freedom humanoid arm for physical AI research in contact rich environments, with high backdrivability, compliance and human scale proportions.

The licensing split is what makes it unusual. The hardware repository carries STL files, STEP files and Fusion 360 assemblies under CERN-OHL-S-2.0, an open hardware licence. Software is Apache-2.0 across URDF and xacro descriptions, a CAN library for the motor bus, ROS 2 nodes, teleoperation with unilateral and bilateral control, Isaac Lab and MuJoCo support, a dataset format with recording tools, and Dora dataflow nodes for collection and inference.

The project also defines OpenArm Cell, a standardised rig with unified background, lighting and camera placement so results compare between labs. That targets a real problem in manipulation research, where policies trained on one bench rarely transfer and the numbers are not comparable. Simulation covers Isaac Lab and Isaac Sim, the usual pairing for sim to real work, and the open question is whether independent groups publish results on the Cell standard.

Source: GitHub.

Arm organises the physical AI stack and names six capability levels

Arm released Arm Total Design for Physical AI, a programme now counting more than 80 companies, according to The Robot Report. Named members include AWS, Hugging Face, Liquid AI, NXP, QNX, Siemens, Unitree Robotics and Qwen. The goal is interoperability earlier in development, spanning software stacks, AI models, sensors, compute hardware, virtual platforms and digital twins, so partners can validate combinations before silicon and integration work is finished.

Arm also introduced a Robotics Capability Framework with six levels, zero to five, inspired by SAE International’s levels of driving automation. Dermot O’Driscoll, Arm’s vice president of go to market and customer solutions for physical AI, described the range as running from robots that react to stimuli under fixed rules, up to robots whose behaviour evolves as they learn and self optimise. Arm says it wants feedback and is open to changing the levels.

Levels of autonomy are politically loaded, but procurement language is the practical payoff for a developer team, alongside a mapping between automotive and robot safety practice: Arm notes that every vehicle on the road already computes on an Arm chip.

Source: The Robot Report.

AthenaZero throws a ball at over 30 m/s

arXiv:2609.19194, submitted on 15 September and the cover article of the September 2026 issue of Science Robotics, describes AthenaZero, a bimanual manipulator designed to cut inertia without giving up control authority. Quasi direct drive actuation combined with transmission remotisation moves mass away from the moving links. Effective endpoint mass is reported as comparable to a human limb, roughly an order of magnitude below conventional manipulators, with torque transparency following from the same design.

The demonstrations are baseball tasks: throwing, catching and batting, with throws exceeding 30 m/s and catching and batting exceeding 14 m/s across a short 7.3 m distance, plus robot to robot catch and human to robot variations. That needs perception and control latency at millisecond timescales rather than pose to pose planning. The claim worth testing is that low inertia, not faster planning, is what enables dynamic contact. The paper is by Andrew S. Morgan and 25 co-authors.

Source: arXiv.

Native ROS 2 on Zephyr, with no DDS agent in the middle

Javier Blanco-Romero, posting as fj-blanco, described running the standard ROS 2 C client stack directly on an ESP32-S3 over Wi Fi, with Zephyr as the RTOS. The chain is rclc to rcl to rmw_cyclonedds_c to Cyclone DDS to Zephyr, and the board joins the DDS/RTPS network itself. The work spans two repositories, ros2_zephyr and rmw_cyclonedds_c.

The author tested the board against stock ROS 2 Lyrical with rmw_cyclonedds_cpp, in both publisher and subscriber directions, and states the limits plainly: fixed size messages and best effort QoS only. Reliable QoS, transient local durability, ROS graph support and DDS Security are on the roadmap, and Micro XRCE-DDS plus rmw_zenoh_pico on Zephyr are being explored for comparison on the same hardware.

The usual embedded ROS 2 path adds a process, a serial link and a protocol translation step between firmware and the system, each one a place to debug. A microcontroller that is a first class DDS participant removes them, leaving footprint and QoS coverage as the open question. Background: what ROS 2 is.

Source: Open Robotics Discourse, author’s notes.

AssemblyGrid v1 benchmarks multi-robot production

arXiv:2609.16075, submitted on 13 September by Fouad Bahrpeyma, David Heik and Dirk Reichelt, introduces AssemblyGrid v1, a reproducible benchmark for repeated multi-robot production under decentralised control. It puts things usually studied separately into one task level formulation: process progression, material transfer between robots, temporary coalitions, productive concurrency and geometry dependent feasibility, all under bounded local information.

The release provides nine official scenarios across Flow, Coalition and Concurrency families at three difficulty levels. Success measures are defined independently of any learning reward or solution method, so learning and non learning methods attack the same problem, and the experiments cover a centralised reference, structured decentralised controllers and MARL methods including IPPO, MAPPO and QMIX. The repository and an interactive demo space are public. Most fleet level claims come from a single integrator’s deployment, so a shared benchmark is a useful counterweight.

Source: arXiv, Open Robotics Discourse.

Quick Bites

• The ROS 2 Kilted Kaiju sync of 17 September brought 16 new packages and 235 updates from 52 maintainers, including robstride-driver and robstride-ros2-control, a ROS 2 control hardware interface for RobStride actuators. ros2_control moved from 5.17.0 to 5.18.1 and behaviortree_cpp from 4.9.0 to 4.10.0. ROS Discourse, rosdistro tag
• An instructor ported the arduinobot manipulator workspace from Humble to Jazzy and wrote up what broke: the ros2 CLI parses URDF parameter values as YAML, so an XML comment containing a colon raises ScannerError: mapping values are not allowed here, and package.xml installs nothing, meaning an install(DIRECTORY urdf ...) rule in CMakeLists.txt is what actually ships the folder. ROS Discourse, repository
• Robotics companies raised $4.87 billion in disclosed funding across 162 tracked transactions in August 2026, per The Robot Report, with China at about $2.44 billion and the United States at $1.27 billion. Unitree Robotics led with a $905 million IPO and XPENG Robotics with a $900 million Series A, while humanoid companies took 19.4% of the disclosed total. The Robot Report
• SMACC2’s PX4 client library has a reference demo that autonomously flies a drone in Gazebo using the PX4 flight stack, a worked example for anyone wiring hierarchical state machines to a flight controller. ROS Discourse, SMACC2