Digit 5, OpenRobOps, Gazebo armhf cutoff

A humanoid built around safety certification, an open source fleet manager for multi-vendor sites, and a build target that reaches end of life next week.

Agility’s Digit 5 targets barrier-free factory work

Agility Robotics unveiled Digit 5 on 15 September, the fifth generation of its humanoid and the first the company says is engineered for “cooperatively safe” work, meaning it can operate near people without the physical barriers traditional industrial automation needs.

The hardware was redesigned for industrial jobs rather than dynamic walking. Digit 5 weighs 284 lb (129 kg), stands 5 ft 11 in (1.81 m), reaches up to 7.2 ft (2.2 m), up from Digit 4’s 5.5 ft, and repeatedly lifts 50 lb (22.7 kg), a 40% payload increase driven by a new leg design with Agility’s cycloidal actuators. The backward, bird-style knee that defined Cassie and earlier Digits is gone: it suited walking and running, while the new one is optimised for lifting and getting up off the ground. A 90-minute battery now charges in nine minutes, a 10:1 run-to-charge ratio against Digit 4’s 2:1, which is how Agility claims more than 20 hours of productive work in a 24-hour day.

Safety is the part that decides deployment: proprietary human detection across multiple sensor types, an independent safety controller that enforces the response, and visual and auditory cues that signal intent to nearby workers. When a person approaches, the robot puts down what it carries and, if needed, sits in a statically stable posture and turns its motors off. Agility states it is the first launch partner for NVIDIA’s Halos for Robotics platform, pairing its system with NVIDIA IGX Thor compute and Halos Core, a stack NVIDIA announced on 22 June. Per Agility, Digit was also the first humanoid to pass an independent field evaluation on a customer production line against OSHA-administered safety standards, and the company contributes to ANSI/A3 TR R15.108 and ISO 25785-1, the emerging humanoid safety standards.

Two interface choices stand out. End effectors are swappable through ISO-standard mounting flanges, and Agility is staying with a two-finger parallel-jaw gripper rather than a five-finger hand. Co-founder Jonathan Hurst told The Robot Report that no hand on the market can pick up 25 kg bins.

The business numbers need caveats. Agility reported more than $300 million in multi-year Digit 5 orders as of May 2026, subject to contractual milestones, alongside a SPAC merger valuing it at about $2.5 billion. IEEE Spectrum, working from Agility’s SEC filings, reports launch bills of materials likely between $150,000 and $200,000 per robot, a target under $50,000 at 10,000 units a year, and a service offering sketched at roughly $8,500 per month. Those are illustrative filing estimates, not prices. Early access is expected in the first half of 2027 and general availability by the end of 2027, with the EU and UK first outside North America.

Sources: Agility Robotics, IEEE Spectrum, The Robot Report.

InOrbit open-sources a fleet operations stack

InOrbit.AI released OpenRobOps on 15 September under Apache 2.0, describing it as a production-grade open foundation for robot operations and fleet management, and the first reference implementation of ISO 21423, the upcoming standard for communication and data exchange between industrial mobile robots and fleet managers.

The repositories are real: OpenRobOps/oro was created in February 2026 and last pushed on 15 September, alongside a Flatland-based simulation and Open-RMF adapters. The feature list is the checklist every AMR vendor rebuilds in-house: telemetry ingest for unreliable networks, spatial tracking with a robot locking mechanism for safe teleoperation, configuration as code so metrics and alert rules live in Git and CI, automated anomaly detection with remediation or escalation, and native Open-RMF adapters for multi-fleet traffic management.

InOrbit frames the build trap as the problem: OEMs spend years maintaining back ends, databases and dashboards instead of their core robotics. Steve Cousins, executive director of the Stanford Robotics Center and former Willow Garage CEO, compares the contribution to the role ROS played for device drivers.

Sources: The Robot Report, openrobops.org.

Gazebo drops 32-bit ARM binaries

The Gazebo team will stop producing armhf (32-bit ARM hard-float) Debian packages. No new armhf packages will be published after Monday 21 September 2026, and armhf is now end of life across all supported releases: Fortress, Harmonic, Ionic and Jetty. Existing packages stay on packages.osrfoundation.org but receive no updates.

If you are weighing simulation options while Gazebo narrows its platform support, what NVIDIA Isaac Sim is and when to use it covers the other stack most robotics teams run.

The reasons hold for most users: the ARM platforms Gazebo runs on, including Raspberry Pi OS, NVIDIA Jetson and ARM cloud instances, ship 64-bit images, new ARM Cortex cores have been 64-bit only since 2023, dependencies such as DART do not support 32-bit targets, and official ROS 2 binaries are amd64 and arm64 only. armhf builds also consume buildfarm capacity and delay releases when builders fail. Building from source may still work, but it is neither supported nor tested.

This matters if your CI images or Dockerfiles pin 32-bit ARM bases, which older tutorials often do. Audit them now and move to arm64, and raise anything unaccounted for on the tracking issue.

Sources: ROS Discourse announcement, release-tools issue 1548.

World-action models, mapped

A 19-page survey submitted to arXiv on 13 September organises World-Action Models, systems that couple future world prediction with executable action generation, and separates them from world models, model-based reinforcement learning, action-conditioned video generation and reactive VLA policies. Its sixteen authors, among them Ivan Laptev, Danfei Xu, Giuseppe Loianno and Eric Xing, build a taxonomy across representations, transition modelling, action interfaces, architectures, training pipelines, data modalities and scaling, then review manipulation, navigation and driving applications with the datasets and metrics used to evaluate them.

The open problems are the useful part if you are choosing an architecture: action alignment, world-action factorisation, spatial and multi-view consistency, long-horizon memory, neural simulation inside closed-loop policy learning, and efficient inference. A map of a research area rather than a result, and a useful one next to our explainer on why VLA models matter.

Source: arXiv:2609.16074, project page.

A benchmark for multi-robot production

AssemblyGrid v1, submitted on 13 September, is a reproducible benchmark for repeated multi-robot production under decentralised control. It puts process progression, material routing, resource assignment, temporary robot coalitions, concurrent execution and geometry-dependent feasibility into one task-level formulation, with each robot acting from bounded local information. Workloads come in Flow, Coalition and Concurrency families with three scenario levels each, and success measures are independent of any learning reward, so learning and non-learning methods solve the same problems.

The authors tested it against a centralised reference, structured decentralised controllers and multi-agent RL methods including IPPO, MAPPO and QMIX, and report that decentralised policies can learn effective production behaviour from local observations. The paper says the implementation will be released publicly; the repository it points to was not available when I checked, so treat the results as paper-only for now.

Source: arXiv:2609.16075.

Quick Bites

• ROS PMC minutes for 15 September record a passed vote promoting RHEL to Tier 1 on Rolling and demoting Fedora to Tier 3, a REP 158 vote still in progress with the Gazebo PMC included, and new committer scopes covering geometry, tf2, launch, rclpy and the rosidl and rmw stack. No PMC meeting next week because of ROSCon. ROS Discourse
• Lyrical packages sync to the main apt repo on 17 September, with 336 packages queued and 256 regressions the release manager expects to clear as the buildfarm catches up. rosdistro release PRs are held until then. ROS Discourse
• University of Manchester researchers retrained NVIDIA’s Earth-2 CorrDiff downscaling model for air pollution, producing a UK-wide model at 2 to 3 square kilometre resolution from a year of hourly simulated data, trained in two days on one eight-GPU node of Isambard-AI and runnable on a DGX Spark. The transferable idea for robotics is the workflow: learn a downscaling model on simulated fields, then infer on small hardware, the gap that also limits environment models built from sparse sensors. NVIDIA
• A contributed article on The Robot Report argues that electronic skin mounted at the contact surface cannot survive high-duty-cycle robotic fingers and makes the case for ultrasound sensing below the outer skin, reading contact through time-of-flight, reflection amplitude, attenuation and phase. The authors work at UltraSense, so read the durability and resolution claims as vendor claims. The Robot Report