Universal Robots Gen 7, Unitree UnifoLM

Two platform releases lead today: an industrial cobot line rebuilt around cameras at the wrist, and open humanoid foundation model weights from Unitree.

Universal Robots unveils Gen 7

At IMTS in Chicago, Universal Robots introduced Gen 7, a redesign spanning the tool flange, controller, teach pendant and software. Three arms join the lineup, named for payload in kilograms and reach in millimetres: UR10g-1750 (8 kg, up to 10 kg extended, 1750 mm reach), UR17g-1300 (15 kg, up to 17.5 kg, 1300 mm) and UR18g-950 (18 kg, 950 mm). UR lists 5 m/s maximum tool centre point speed for the first two and 4 m/s for the third.

The developer-relevant part is the g-Series tool flange, which carries 24 or 48 V power (5 A peak, 3 A continuous), 1 Gb Ethernet for cameras and sensors, and M8 I/O with digital I/O, analog input and RS-485, so vision can run at the wrist without extra cabling.

UR states the CB7 Core controller delivers 40% more compute power in a 30% smaller footprint than the previous generation, with 3x 1 GbE ports and 16 configurable digital I/O. UR’s own Gen 7 product page lists a 20% footprint reduction for CB7 Core, so treat the size figure as approximate. Safety is certified as PLd, Category 3 under ISO 13849-1, with the g-Series arms TÜV certified to ISO 10218-1 and UL1740. PolyScope X advertises open APIs and ROS 2 communication so external IPCs can host AI processing, and Inbolt’s 3D vision now runs natively on the controller instead of a separate vision PC. Watch whether that ROS 2 path matures into something teams can build on, or stays a bridge for partner software. Background: what ROS 2 is.

Source: Universal Robots press release and the Gen 7 product page.

Unitree opens UnifoLM-WLA-1.0 weights

Unitree published the first parts of UnifoLM-WLA-1.0, its next-generation humanoid foundation model. Per the project page, it has 6B parameters, was trained on roughly 2,500 hours of real-robot data, and uses one shared set of weights across 64 tasks, split into 54 tabletop and 10 whole-body manipulation tasks, covering two-finger grippers and several five-finger dexterous hands.

Architecturally it is a two-stage stack. The embodied reasoner, UnifoLM-ER-1, is built on Qwen3-VL-4B and trained on more than 5 million samples spanning point prediction, object detection, multi-image reasoning, 2D trajectory and 3D detection. UnifoLM-ER-Flow adds discrete action tokens from residual vector quantization over end-effector poses, end-effector joints and lower-body joints, plus mask tokens for predicted future dynamic regions. The full WLA model adds an MMDiT action expert. Unitree claims ER-1 leads open-source models on seven of 16 multimodal benchmarks, so read those results as vendor figures. Background: why VLA models matter.

What shipped on 11 September is weights for UnifoLM-ER-1 and UnifoLM-ER-Flow on Hugging Face, plus the UniBot-V1 Challenge dataset. Post-training code, WLA-Base weights and the Unitree-WBT and manipulation datasets are still pending, and the repository has no tagged releases yet. Real-robot evaluations shown are on the Unitree G1.

Sources: repository, project page, Hugging Face model.

OpenArm ships an open 7DOF humanoid arm

OpenArm, from Enactic, is an open-source 7 degree-of-freedom humanoid arm aimed at contact-rich physical AI research, built for backdrivability and compliance. The project spans ten repositories, including CAD in STL, STEP and Fusion 360 under CERN-OHL-S-2.0, URDF and xacro descriptions, a CAN motor control library, ROS 2 integration, bilateral teleoperation, an Isaac Lab environment, MuJoCo assets and a dataset format with Python API. The software is Apache-2.0.

The most useful idea is OpenArm Cell, a standardised evaluation environment with fixed background, lighting and camera placement, so results from different labs can be compared under consistent conditions. For manipulation and imitation learning work it pairs real hardware with simulation parity across Isaac Lab and MuJoCo and ROS 2 or MoveIt 2 integration in one stack. The repository shows about 3.1k stars and 354 forks. Simulation background: what NVIDIA Isaac Sim is.

Source: OpenArm on GitHub.

Mecka AI reportedly nears $500M for robot data

TechCrunch reported that Mecka AI, which collects human motion data to train humanoid robots, is nearing a round led by Sequoia Capital at a valuation of about $500 million, citing two unnamed people. The report says the terms are not final and that neither Mecka nor Sequoia commented, so treat the valuation as reported rather than confirmed. The company raised $60 million three months earlier, led by Framework Ventures. TechCrunch also reported that XDOF is nearing a round at a $1.2 billion valuation.

The mechanism is egocentric data collection, paying people to record everyday tasks with body sensors and smartphones, the same supply-side bet Scale AI made for language models. The takeaway is that capital currently sees scarcity in robot training data rather than model architecture, so expect more third-party datasets with their own licensing and quality caveats. See also the robot data bottleneck.

Source: TechCrunch.

ShieldVLA aligns VLA policies for safety

A new paper proposes ShieldVLA, a safety-aligned fine-tuning framework for vision-language-action models built on Hamilton-Jacobi reachability. Instead of the usual Lagrangian soft penalties, the authors learn a model-free approximation of the reachability value function from visual observations, then use the safety critic to gate policy optimisation, separating reward maximisation inside the safe region from recovery near unsafe states. Per-step cost labels are replaced by rubric-based VLM safety scores.

The authors report an average 57% reduction in cumulative safety cost and a 0.13 absolute improvement in task success rate over SafeVLA across five navigation and manipulation benchmarks and multiple VLA backbones. This is simulation and benchmark evidence, so the open question is how the reachability critic behaves on hardware where the visual distribution shifts.

Source: arXiv:2609.13231.

Self-evolving humanoids, and what is missing

A survey from six authors examines what changes when a self-improving agent has a body. It defines a deployed humanoid as a state tuple of policy, perception, memory, workflow and body, updated by an evolution operator in a slow outer loop, and organises the literature into four mechanisms ordered by increasing autonomy: self-learning, self-adaptation, self-optimisation and self-generation. Safety is a design dimension of that operator, with admissible evolution constrained by a world-model verification gate inside a human-oversight envelope.

The sharpest point is evaluation: progress should be tracked as trajectories rather than compared at fixed checkpoints, and the authors state that no benchmark exists for self-evolving humanoids. It is 30 pages and under consideration for publication, so read it as a map of a research area, not a result.

Source: arXiv:2609.13236.

ROS 2 release train moves

A Humble Hawksbill patch release on 14 September brought 77 new packages and 687 updates, tagged humble/2026-09-14. Additions include moveit_py 2.5.10, a robstride actuator driver, Elite Robots driver packages, Kangaroo and TIAGo Dual bringups, and a MuJoCo 3D LiDAR package. Separately, the Kilted distribution is preparing a sync for 16 September with 248 packages queued and rosdistro pull requests held from 15 September. If you pin to a distro, read the release threads before upgrading this week.

Sources: Humble patch release and Kilted sync.

Quick Bites

• NVIDIA and Perplexity brought Portable Computer, Perplexity’s local agent, to Windows PCs. It needs a GeForce RTX or RTX PRO GPU with 24 GB or more VRAM and ships with a local model NVIDIA names as Qwen 3.8 27B. That 24 GB envelope is the one many robotics teams already use for edge perception and local policies on RTX hardware and DGX Spark. NVIDIA
• A ROS Discourse proposal argues that any robot policy resilience number should be reported next to an unperturbed control arm run on identical seeds. The author measured 19 perturbation families and 8 policy adapters in simulation and found the unperturbed arm fails at a non-zero, family-dependent rate, and notes there are no hardware results yet. ROS Discourse
• IEEE Spectrum’s Video Friday covered monkey-bar traversal work from ETH Zurich’s Robotic Systems Lab, where a humanoid must perceive thin overhanging geometry and land safely, and marked the anniversary of the first documented use of robots for urban search and rescue. IEEE Spectrum