Digit 5, SoftBank RAI, JOY microgravity

A humanoid built around a safety argument rather than a payload number, a research institute changing hands, a free-flying robot surviving zero gravity, and robot security work that moved into the model itself.

Agility’s Digit 5 is a safety argument with a robot around it

Agility Robotics unveiled Digit 5 on 15 September, calling it its first humanoid engineered for cooperatively safe work at scale: able to work beside people without the barriers that surround conventional industrial automation. IEEE Spectrum frames the trade-off, since power, safety and battery life normally fight each other, and Spectrum reports that Agility’s answer is to make falling onto a person something Digit avoids by construction, stopping or seating itself instead.

The specifications are Agility’s own. Digit 5 lifts up to 50 lb (22.7 kg), a 40% payload increase built on a new leg design with cycloidal actuators, reaches 2.2 m, and weighs 129 kg at 1.81 m. Its 90 minute battery charges in 9 minutes, stated as a 10:1 run-to-charge ratio against 2:1 on Digit 4, for more than 20 hours of work in a 24 hour day. Three safety mechanisms are claimed: human detection from proprietary AI plus multiple sensors, visual and audible motion-intent cues, and an independent safety controller that forces a response when someone is inside an unsafe distance.

The infrastructure detail matters more for developers than the payload. Digit 5 is the first launch partner for NVIDIA’s Halos for Robotics platform with IGX Thor, and Agility is a project leader on ANSI/A3 TR R15.108 and a contributor to ISO 25785-1, the emerging humanoid safety standard. Swappable end-effectors on ISO-standard flanges and Agility Arc integration across WMS, WES and MES fill out the platform. Digit 4 has logged more than 65,000 operational hours across GXO, Schaeffler, Amazon and Toyota Motor Manufacturing Canada, and Agility says it held over $300 million in multi-year orders as of May 2026, subject to contractual milestones, with early access in the first half of 2027 and general availability by year end. All safety figures are vendor reported, and independent evaluation of the Halos-based architecture is what to watch.

SoftBank agrees to acquire the Robotics and AI Institute

The Robot Report reported on 18 September that SoftBank Group has agreed to acquire the Robotics and AI Institute (RAI) from Hyundai Motor Group, citing multiple sources, with the deal under review by the Committee on Foreign Investment in the United States. Terms are unknown and neither company confirmed, with RAI saying it had nothing to share. Treat this as reported, not announced.

The lineage is the story. Marc Raibert founded both Boston Dynamics and RAI, and RAI was spun out of Hyundai’s Boston Dynamics acquisition in 2022 with an initial investment reported above $400 million, funding that has now ended. Its whole-body learning framework sits behind the acrobatic Atlas behaviours. SoftBank owned Boston Dynamics from 2017 to 2021, bought from Google for a reported $100 million and sold to Hyundai at a $1.1 billion valuation, and it is separately buying ABB’s robotics business for $5.3 billion. The question for research labs is whether RAI keeps its publication and open-source output.

Icarus Robotics tests JOY in microgravity

Icarus Robotics flew its free-flying robot JOY on four parabolic flights between 9 and 11 September, 66 parabolas with roughly 20 seconds of weightlessness each. Payload reports the campaign exercised three capabilities in zero gravity for the first time: manipulation, with the arms tracking commands and recovering from disturbances as the preflight models predicted; state estimation, holding position and velocity through each weightless period and both transitions; and flight control, where a settling period after release introduced variance the team does not expect on station.

JOY has two in-house arms with seven degrees of freedom and adaptive pinch grippers. Teleoperation was not possible in flight, so Icarus recorded trajectories and set points on the ground, replayed them in microgravity, and compared the runs. The arms were the hardest subsystem to model in simulation and held up best, a familiar sim-to-real result for contact dynamics. Hardware goes to NASA in January, then to the ISS in 2027 under the Voyager mission agreement, starting teleoperated and moving toward autonomy. Icarus raised a $6.1 million seed round in 2025, and Payload reports the mission was upgraded from a demo to a full operating payload.

NVIDIA’s MLPerf v6.1 run includes an edge result aimed at robots

NVIDIA published MLPerf Inference v6.1 results on 16 September. Vera Rubin NVL72 made a first preview submission with up to 3.7x higher throughput than GB300 NVL72 on Qwen3-VL using vLLM and Dynamo, and up to 2.5x on DeepSeek-R1 with TensorRT-LLM. A GB300 NVL72 run scaled from 72 to 288 GPUs across four racks at 99% scaling efficiency, and software changes alone lifted Qwen3-VL by up to 1.6x over the v6.0 results.

The robotics line is quieter: Jetson AGX Thor results using TensorRT Edge-LLM on the new Edge-Agentic benchmark with Qwen3.6-27B. Robots run a policy or a vision-language model on a power budget, not rack scale inference, and multi-step agentic work is where embodied stacks are heading. All figures are vendor submissions and Vera Rubin is a preview.

InOrbit releases an ISO 21423 reference implementation

InOrbit.AI released OpenRobOps on 15 September, an open source robot operations and fleet management platform it calls the first ISO 21423 reference implementation. That standard defines how autonomous mobile robots and fleet managers talk to each other, and the Open Robotics weekly update notes it supports ROS 2 and Open-RMF out of the box.

The target is the build trap: every OEM that sells a fleet maintains a custom backend, database and dashboard. InOrbit’s announcement lists telemetry ingest for unreliable networks, spatial tracking and fleet observability as the foundation. Check the licence and standards coverage before betting a production fleet on it, and note the conformance claim is the vendor’s.

Robot security moves from the network into the model

A sponsored article from VicOne in IEEE Spectrum summarises work worth reading through its primary sources. On vision-language-action models, BadVLA (May 2025) plants a backdoor using objective-decoupled optimisation with near-100% attack success while preserving clean-task accuracy, surviving task transfer and fine-tuning. FreezeVLA (September 2025) generates adversarial images that freeze VLA models so they ignore later instructions, reporting 76.2% average attack success across three models and four benchmarks, transferring across prompts. Both live in the training pipeline or the visual input, not the network stack.

The network layer has its own entry. UniPwn documents a Unitree Bluetooth Low Energy command injection chain with four assigned CVEs, CVE-2025-35027, CVE-2025-60017, CVE-2025-60250 and CVE-2025-60251, and arXiv:2509.14139 presents a security assessment of the Unitree G1. The repository notes newer Unitree platforms appear to share the Go2 codebase, so impact is likely wider than the tested robot.

Clean-input evaluation is not enough: a policy can carry a conditional behaviour that appears only with a trigger, which makes model supply chain and trigger testing part of safety engineering. VicOne has published a free Isaac Sim extension for testing adversarial visual inputs against robots. For what that simulator is and where it fits, see what NVIDIA Isaac Sim is. Background: why VLA models matter.

Quick Bites

• The ROS 2 Lyrical sync of 18 September added 46 packages and 315 updates, including Navigation2 1.5.1, realsense2-camera 4.58.4, mujoco-3d-lidar, DepthAI 3.10.0 and the new lgdx_rplidar_c1 driver. ROS Discourse, rosdistro tag
• LGDXRobot2, an open source mecanum wheel robot, now supports ROS 2 Lyrical with a refactored C++20 controller node, an ICM-20948 9-DoF IMU on the redesigned board, Nav2 integration and a Docker image with browser GUI access. ROS Discourse, GitHub
• Universal Robots launched Gen 7 at IMTS on 14 September: the PolyScope X operating system with a re-engineered control box, teach pendant and tool flange, plus three larger arms named for payload and reach, UR10g-1750, UR17g-1300 and UR18g-950. The Robot Report
• Neptune Medical received FDA clearance for its Triton 1 robotic colonoscopy system. The company reports its first-in-human CARE 1 study met primary endpoints with no adverse events, 100% cecal intubation and a 54.2% adenoma detection rate against an industry threshold of 35%. The Robot Report