Section 232, Starship fault, BMW policies

A trade proclamation with no tariffs in it yet, a delivery robot that misread a traffic light, and two very different bets on where humanoid competence comes from.

Washington opens robotics import negotiations

President Trump signed a Section 232 proclamation on 28 October covering robotics, industrial machinery and their parts and components, after the Commerce Department found that imports of those goods threaten to impair national security (White House).

Nothing is taxed yet. The proclamation directs the Commerce Secretary and the US Trade Representative to negotiate agreements with trading partners, with tariffs, quotas, fees or a domestic production incentive left as later options. The first status update is due within 180 days.

The finding names autonomous, humanoid and quadruped robots and autonomous ground vehicles as military support, and cites roughly $7.1 trillion in manufactured goods shipments in 2024.

For robot builders the exposure sits in components, which the proclamation covers. Robert Little, formerly chief of robotics strategy at Novanta, told The Robot Report that sensors, actuators, precision gears and encoders are the weak part of the domestic supply base, while Boston Dynamics policy lead Brendan Schulman described six months of negotiations before any trade measure lands (The Robot Report).

The practical read: landed cost and lead time on imported drives, gears and sensors now carry policy risk, while order volumes are still climbing.

Starship robot runs a red light and hits a bus

On Monday evening a Bolt delivery robot built by Starship entered the Luise and Tehnika crossing in central Tallinn on a red light and collided with a bus, which braked hard. Tallinn Transport says up to nine standing passengers were injured (ERR).

Starship’s head of operations, Pepe Aaviksoo, gave the cause as perception under bad conditions: it was dark, an illuminated advertisement sat behind the traffic lights, pedestrian and vehicle signals were in the same line of sight, and the robot’s light sensitivity sensor was not working as it should and had changed for some reason. The company says the fixes will be mostly software (ERR follow-up).

Two details transfer to any sidewalk robot program. Remote takeover exists, but robots are not continuously watched, which Starship calls uneconomic across several thousand machines. And Estonia’s Transport Administration requires a robot to cross only when safe, to give way, and after a crash to stop and have its operator call the Emergency Response Center if injury is visible.

The engineering lesson is sensor health monitoring across lighting conditions, plus logs good enough to reconstruct what the perception stack saw. rosbag2 is where most teams start.

BMW builds humanoid software at Landshut

BMW’s Landshut parts plant is developing software for AI-supported humanoid robots in component production, complementing humanoid pilots already running at Leipzig in Germany and Spartanburg in the United States (Assembly).

The scope is the unglamorous half of physical AI: open software platforms, generation and processing of training data, simulation and motion planning, and training the systems themselves. The plant combines conventional deterministic programming with models that take in vision, language and robot actions.

Data collection is explicit. The team records real production tasks and test setups with cameras, motion capture suits and data gloves, then converts the demonstrations into policies that can be applied to other tasks. New approaches are tested outside running production, moved into pilot projects, and scaled in stages. BMW frames it as a requirements exercise: physical AI project manager Christoph Jagoda says the target is what production needs, not what is technically possible.

Notably, BMW does not name a robot vendor for Landshut, so the software and data layer is meant to outlive any single platform. The same argument favors a shared data pipeline: demonstrations are only an asset if they can be replayed, relabeled and retrained.

BTick adds breakpoints to ROS 2 behavior trees

A developer released BTick, an open source tool for monitoring, debugging and runtime control of BehaviorTree.CPP applications in ROS 2 (ROS Discourse, core repo).

The feature list covers what Groot does not. BTick discovers multiple trees and nested subtrees, shows live node states, supports breakpoints and step execution, lets you override node status or inject delays, and edits ports and blackboard variables at runtime. It records and replays sessions on a timeline, reports execution timing, includes an XML editor for BehaviorTree.CPP v4, speaks its own protocol plus the Groot2 protocol with some Pro features, and reads Nav2’s behavior tree logging topic.

The caveat is maturity. The core repository was created on 12 October 2026, is written in C++ and had two stars when checked, so treat it as a promising early tool rather than infrastructure. Still, runtime state injection is what makes behavior trees testable: you can force a recovery branch to fail without unplugging a sensor. Teams debugging Nav2 trees have wanted that for a while.

AGIBOT proposes a productivity metric for robots

AGIBOT used its Innovation Day in Shanghai to propose Worken, a framework for measuring how much useful work deployed robots actually deliver, alongside a hardware platform, an updated AGILE motion framework and a preview of its GO-3 embodied foundation model (The Robot Report).

The formula is active robots multiplied by autonomous working hours multiplied by an output efficiency ratio, where the third term compares acceptable output against a reference such as a human worker under established standards. The motivation is that unit counts say nothing about uptime, interventions or rework.

AGIBOT says its 20,000th robot left the line last month, and IDC reported it led global humanoid shipments in the first half of 2026.

Worken is a vendor proposal, not an audited standard, and the output efficiency ratio is the hard part: defining acceptable work per task is exactly where deployments differ. The useful idea for developers is measuring autonomy rate and intervention rate as first class numbers, the same quantities that decide whether a policy is deployable or merely demonstrable.

Artificer scores task progress with 4B parameters

Nua published Artificer, a 4B parameter model that estimates task completion and progress from video, intended as feedback during training, evaluation and deployment (Nua).

The problem it targets is real: a policy can grasp the right object and place it in the wrong bin, and judging that from video usually means human review or a handcrafted success check. Nua evaluated Artificer on RoboReward, comparing task completion scores against reference scores, and on ArmNet268 for success detection, using a fixed threshold and balanced accuracy.

The company says the model reaches frontier level on those benchmarks while being cheaper and faster than large general models. The page shows the comparison only as charts, with no numeric table, and there is no independent replication yet, so treat that claim as the vendor’s.

The deployment use is the interesting one: a controller that can tell a partial attempt from a failed one and decide to continue, retry or move on. That is a different job from training a VLA policy, and it is the piece most manipulation stacks still do with thresholds.

Quick Bites

• A GNSS vendor published field logs from RTK receivers under roadside trees and rain: three of four receiver groups held 100 percent fixed over long baselines, one dropped to 83.3 and 88.5 percent, and the author notes that fix ratio is not accuracy without ground truth (ROS Discourse).
• Symbotic’s largest customer supplied 90.5 percent of revenue in the quarter ended 27 June 2026, on $720.8 million of revenue, and Walmart’s contracts include engineer embedding and escrow rights (The Robot Report).
• NVIDIA published a walkthrough of using frontier AI agents inside Omniverse to generate and iterate simulations, the workflow that precedes policy training for many teams (NVIDIA).
• Boston Dynamics employees are petitioning the company to cancel a robot dog sale to ICE, alongside the agency’s plan (Ars Technica).