Robot imports become a security question

Robot imports are now a national security file in Washington, Agility became the first pure play humanoid company with a ticker, and two open-weight models attacked the same data bottleneck from opposite directions. Also this week: a force command channel for humanoid whole body control, and a Rust planner that batches inverse kinematics across a GPU.

Washington opens a tariff path for imported robots

President Trump signed a proclamation on 28 October finding that imports of robotics, industrial machinery and their components threaten national security, and directing the Commerce Department and the US Trade Representative to negotiate agreements with trading partners (proclamation, Ars Technica).

No tariff applies today. The agencies have 180 days to report back, and the proclamation holds out tariffs, duties, quotas and fees, plus a programme to subsidise domestic production, if talks fail or an agreement is not carried out.

The findings are the interesting part. The proclamation puts robotics and industrial machinery behind roughly 7.1 trillion dollars in manufactured goods shipments in 2024 and about 86 billion dollars of US equipment production, and states plainly that the country relies on imports for most types of robotics and does not produce enough to meet national security needs.

Scope matters. The FCC added foreign produced advanced robotic devices to its Covered List in July 2026, blocking new models of humanoids, quadrupeds and mobile robots without a Defence Department approval, but exempted fixed industrial arms. Sidley, citing the Council on Foreign Relations, notes that Canada, Germany, Italy and Japan supply about half of US robot and machinery imports while China ships 8 percent (Sidley). This proclamation reaches the arms, and Ars notes that FANUC, ABB, Yaskawa and KUKA supply most of the heavy industrial arms bought in the US.

For anyone specifying a cell, the exposure is landed cost and lead time on the arm. A tariff would raise the price of hardware most US integrators already buy, in a year when the International Federation of Robotics counted nearly 38,500 new US industrial robot installations.

Agility starts trading as AGLT

Agility Robotics began trading on Nasdaq on 5 November under AGLT, with warrants as AGLTW, after closing its business combination with Churchill Capital Corp XI (press release). The company says the deal raised about 465 million dollars gross before expenses, including a PIPE led by Foxconn.

Agility reports more than 65,000 hours of Digit operation, more than 100,000 totes moved at the GXO site in Georgia, and two more GXO facilities scheduled for early 2027. It claims more than 300 million dollars of committed multi-year orders for Digit 5, subject to contractual milestones, product features and specifications, a qualifier worth keeping attached. TechCrunch adds the other side: 174,000 dollars of revenue in the first half of 2026, mostly from related parties, against a 362 million dollar loss, with customers paying 500,000 dollars for an on-site proof of concept before a deployment scales (TechCrunch).

Digit 5 is the bet. Agility describes cycloidal actuators built for repeated lifts of up to 50 pounds, interchangeable end effectors, cooperative safety so the robot can work near people rather than in a fenced cell, NVIDIA IGX Thor as the onboard computer, and a Nyobolt battery the company says recharges in nine minutes for more than 20 productive hours in a 24 hour day. Early customer access is targeted for the first half of 2027, general availability for the end of 2027, from a 70,000 square foot RoboFab in Salem, Oregon, designed for up to 10,000 robots a year.

Cooperative safety is the part developers should read closely, since a robot rated to work alongside people changes cell layout, sensor set and the safety case. IGX Thor is the hardware class covered in our Jetson comparison.

Verne V0.5 opens its weights

Verne Robotics released V0.5, an open-weight world action model for dual arm mobile robots, with weights on Hugging Face and code at vernerobotics/v0.5 (research post).

The stack is four layers: an off the shelf vision language model that turns a goal into the next short instruction, the policy that outputs 30 Hz action chunks of end effector poses, gripper commands and a base position, an execution layer that converts chunks into joint trajectories while planning the next one, and a 500 Hz controller. Pretraining used more than 1.7 million demonstrations, mostly teleoperation and handheld gripper data.

Reported results, all the company’s own: 99.3 percent average success across the four LIBERO suites, 99.2 percent on the Long suite, and 81.0 percent clean and 73.4 percent randomised on RoboTwin 2.0 after fine-tuning on 2,500 clean demonstrations. The headline claim is scoped to the leading US open-weight model, and fairly so. On the same chart ABot-M0.5 scores 99.4 percent on LIBERO and PICA-WAM reaches 90.7 and 74.7 on RoboTwin, so V0.5 leads the US group and trails the best published open numbers elsewhere.

The scaling result is the reusable finding. With no fine-tuning, RoboTwin success rose from 23 to 51 percent as pretraining data went from 29 to 100 percent of the corpus, which repeats the argument in our world action model survey: capability tracks data coverage more than architecture novelty. Verne also ships a fine-tuning platform, and claims a new task takes under two hours of demonstrations.

Ctrl-F gives humanoids a force command

A Berkeley team added force to the command channel for humanoid loco-manipulation. Ctrl-F takes a reference motion plus an explicit three dimensional force at each designated contact body, a hand or a foot, and tracks both (arXiv 2611.05507, project page).

Stage one trains the tracker with no objects at all: each episode samples a force command and applies the opposite wrench, so holding the reference pose requires pushing back with exactly that force. Stage two freezes the tracker and trains a planner to pick the force command from object state and proprioception, with an analytical wrench allocation model as teacher and reinforcement learning on object motion to refine it.

On a Unitree G1 with AprilTag object pose, Ctrl-F dragged, pushed, lifted and kicked boxes up to 20 kg, and adapted when load changed mid task. The baseline is a force blind tracker running the same motion clips on the same objects, which the authors report moves nothing in those conditions. Code is listed as coming soon.

The framing is the contribution: motion tracking answers where the body should be, not how hard it should press. Our comparison of pose and force action spaces made the point for hands, and this extends it to legs.

batchplan 0.5 batches IK on any GPU

batchplan, a Rust motion planning library, shipped 0.5, and the releases from 0.3 up added tool paths, pick and place, blended motions, live replanning and servoing to the core library (ROS Discourse, library, ROS bindings).

The design choice worth noting is the backend. Batched inverse kinematics and trajectory optimisation run through wgpu, so the GPU path covers Vulkan, Metal and DX12 without a CUDA dependency, and the same code falls back to CPU. The planner does batched IK, RRT-Connect with SIMD collision checks and B-spline paths, optionally refined with L-BFGS and timed within velocity, acceleration, jerk and hand speed limits. The ROS 2 bindings add Cartesian paths and blended sequences on Lyrical and Jazzy, with a MoveIt plugin and a Panda example that tries several grasps in one request.

It is a young project with one maintainer and no benchmark against MoveIt’s planners, and batch planning pays off with many candidate grasps or many robots, not one arm following one path.

Quick Bites

• DC 0.2, an open source data collection framework for ROS 2, now runs on Jazzy, Lyrical and Rolling: it validates each record against a JSON schema and ships it to PostgreSQL or object storage (ROS Discourse).
• LiberAI reports 36.0 percent success over 600 trials on 60 tasks for Liber-0, a world action model pretrained on 2.2 million hours of human video and glove recorded hand motion; the article is sponsored content, so the numbers are the vendor’s (The Robot Report).
• Ford is putting 2 billion dollars into its Louisville Assembly Plant for a three branch system that builds major vehicle sections in parallel, claiming 40 percent faster assembly than a conventional line (Assembly Magazine).