Cable robots and open robot weights

Four motors, four ceiling anchors and a gripper on lines: that is the week’s most discussed robot, and it costs less than a laptop. The same week brought open robot weights with benchmarks attached, a warehouse robot rebuilt around service time, and a standard for mixed fleets. The common thread is cost of entry, in hardware, model weights and integration.

Stringman hangs a robot from the ceiling for $795

Neufangled Robotics opened orders for Stringman, a cable driven robot that suspends a two finger gripper from lines anchored in the four upper corners of a room. It reached 840 points on Hacker News on one number: 795 dollars for a machine that works across a whole room instead of a fixed radius (Neufangled).

The architecture is the argument. Four motors share the load, so payload scales with combined torque rather than one actuator: up to 2.2 kg near the floor and 1.9 kg average across a 7 by 7 by 3 metre room at the 28 newton line tension limit, net of a 638 gram gripper. Compute is external, a Raspberry Pi Zero 2W in the gripper and Pi 3A+ boards in the anchors, with your own machine doing the planning.

The interfaces matter more than the price. CAD, PCBs and control software are published, the hardware is OSHWA certified under US002788, and the robot runs air gapped on a local network. There is a Python API and an MCP server aimed at agents, controller in the loop MuJoCo simulation, LeRobot compatibility, and datasets on Hugging Face. The company says as few as 50 episodes teach some behaviours, which makes it a cheap testbed for cross embodiment work.

Two caveats. The claim of roughly 500 times the work volume per dollar of a 2,500 dollar research arm is vendor arithmetic with assumptions attached. And cable driven parallel robots are unforgiving in ways serial arms are not: lines must stay in tension, stiffness and payload margin shift as the geometry changes, and the hanging gripper swings unless something cancels it. Stringman answers that with an active swing cancellation loop on the existing motors, and the 28 newton limit doubles as the safety story, since the motors relax and the lines pull loose.

Verne releases V0.5, an open weight world action model

Verne Robotics published the weights of V0.5, a world action model for dual arm mobile robots (Verne). It follows the two system pattern that GR00T N1 and Helix also use: an off the shelf vision language model turns a goal into short instructions, and V0.5 turns each instruction into motion.

The policy predicts future video and actions jointly through a video transformer and an action transformer, built from the Wan2.2 video model and pretrained on more than 1.7 million demonstrations. At runtime it emits 30 Hz action chunks of end effector poses, gripper commands and base position, generates no video, and hands off to a separate C++ controller at 500 Hz on a real time kernel.

The published numbers, from Verne’s own runs with each benchmark’s standard protocol: 99.3 percent average on LIBERO’s four suites, against 96.85 for pi0.5 and 94.15 for pi0. On RoboTwin 2.0, 81.0 percent in clean scenes and 73.4 percent randomized, against 70.7 and 46.0 for pi0.5. The robustness gap is the interesting part, since clutter, textures, lighting and table heights cost V0.5 7.6 points and pi0.5 24.7. The claim is scoped to US companies as of 1 November 2026, with ABot-M0.5 slightly ahead on the LIBERO average at 99.4 but behind on Long at 98.4.

The controller study is worth stealing. Against a stock YAM controller, Verne reports the arm trailing the command by 52 to 60 ms with median error rising from 9 mm to 26 mm at speed, and 4 to 12 ms and 2.2 to 3.5 mm with its own controller over 20 runs. At 30 Hz that lag is nearly two command targets, which breaks real time chunking. Hence end effector space and a tuned loop, as our piece on action spaces argues from the other side. Background on conditioning is in our world action model primer.

Boston Dynamics rebuilds Stretch around uptime

Boston Dynamics announced Stretch Gen 2 on 3 November, designed from field data: it says it analysed more than 100,000 operating hours and tens of millions of boxes to find what took robots offline (Boston Dynamics, The Robot Report).

The headline rate is up to 1,000 cases per hour with multipick, which the company qualifies as highly dependent on case weight and stack neatness. The engineering is more instructive. Boston Dynamics cut 81 percent of the internal cabling, about half a mile, and moved to automotive grade latching connectors sourced through Hyundai’s supply chain. Joints rotate continuously past 360 degrees, each suction cup has its own sensor and valve so multipick does not waste pump power, and a swerve drive swap takes about an hour instead of a day.

Then the claim that matters for anyone running a fleet: a redesigned compute platform, roughly monthly updates, and a recent release that produced a 24 percent increase in unload rate across deployed robots. That is a vendor figure, but it locates the margin in perception of boxes, conveyors and container walls, pushed to hardware already in the field.

ISO 21423 gives mixed fleets a common language

A3 announced publication of ISO 21423:2026, “Robotics, Industrial mobile robots, Communications and interoperability”, which specifies how mobile robots, fleet managers and other equipment exchange information and request actions (A3, ASSEMBLY).

Robots share location, speed, direction, condition, tasking and availability, and the scope runs past robot to robot traffic: machines tended by robots, ERP systems, lifts, doors, alarms, and worker devices so people are coordinated alongside the fleet. VDA 5050, the German automotive industry’s AGV interface, already covered the fleet manager to robot link; this is the wider, international version.

The developer consequence is a conformance target, not a library. The text is paywalled, listed at 227 Swiss francs in the ISO catalogue, so most teams will meet it through vendor documentation. Anyone whose fleet manager was built against the draft should re-read the published requirements, because a DIS era connector is not a conformant one. For sites held together by custom glue around Nav2 or vendor APIs, this is the first document that says what the glue should say.

ZED 5.6 sends camera data straight from GPU memory

Stereolabs released ZED ROS 2 Wrapper 5.6, which supports the rosidl::Buffer fields added to ROS 2 Lyrical and the CUDA buffer backend NVIDIA contributed with the Open Source Robotics Alliance, the mechanism underneath Isaac ROS 5 (ROS Discourse).

The node publishes images, depth, confidence and point clouds directly from GPU memory, so a GPU subscriber skips the copy out, serialize, deserialize, copy back round trip. Interfaces stay standard: sensor_msgs/msg/Image and PointCloud2, no custom types, no NITROS bridge. Subscribers that cannot take CUDA data, including RViz2, ros2 bag, Python nodes and remote machines, get CPU data automatically, selected by general.buffer_backend, which defaults to cuda. Fast DDS and Zenoh both work, the latter at rmw_zenoh_cpp 0.10.5 or newer.

Measured on a Jetson AGX Thor developer kit with a ZED 2i and Isaac ROS 5 on bare metal, Stereolabs reports latency reductions from 10 percent on a compact point cloud to 61 percent on a full resolution 14.7 MB cloud, and about three times lower standard deviation. One trap: intra-process subscribers get CPU data, so disable IPC for GPU buffers, a tradeoff covered in our guide to intra process composition.

What one tour of Unitree’s factory showed

Josef Chen published a field report from a tour of Unitree’s Hangzhou plant, and the observation that carried it to 132 points on Hacker News was that the robots are not built by robots (josef.cn). In the actuator building he saw one or two automated lines and a floor of people at conventional coil winding machines, with photography not permitted.

The numbers he pulls from Unitree’s filings explain why. Labour was 8 percent of the cost of goods in 2025 against 81.7 percent materials, and the company expects its new motor line to cut labour per motor by 57 percent while cutting total cost per robot by 1.7 percent. It made almost 32,000 robots that year, including about 26,000 quadrupeds, and three quarters of humanoid revenue in the first nine months of 2025 came from research and education, under 5 percent from factories.

One observer, one plant, and figures read from Chinese language filings. The conclusion for developers still holds: the parts bill sets the price, not assembly labour, which is why Unitree automates joint modules first. They stay put while the robot around them changes every year.

Quick Bites

• Gazebo Harmonic and Fortress lose macOS support on 10 January 2027, when Homebrew disables the freeimage formula over open CVEs; Ubuntu support continues to the documented end of life dates (ROS Discourse).
• MABEL, a 56 degree of freedom bimanual mobile robot from NYU and the University of Toronto, is fully open: hardware under CERN-OHL-S-2.0, software under Apache-2.0, two 7 DOF OpenArm arms, two 17 DOF ORCA hands, and teleoperation from one Apple Vision Pro. The authors put the Essential build near 8,056 dollars in parts (ROS Discourse, repository).
• Physicl raised 10 million dollars in seed funding led by Inovia Capital for simulation tooling aimed at post-training on hard edge cases, with Studio shipping 10,000 3D environments and 100,000 assets for synthetic data (TechCrunch).