White House Calls Robot Imports a Threat, AI Models Get Twice as Good at Robots Every 54 Days

On 28 October the White House declared robot and machinery imports a national security risk, whatever country they come from, and imposed no tariff. The same day, the trade group A3 reported that North American robot orders rose 24.8% in the third quarter, and many of those robots are imports. Tesla said it has started building Optimus and wants 20,000 a week by the end of 2027, while Agility's first customer, GXO, said they have not paid for themselves yet. On the research side, frontier models are getting better at driving a real robot arm quickly, and three papers argue that a robot policy does not need to generate video while it runs.

The Section 232 proclamation on robots

The proclamation covers robots, industrial machinery and their parts, and it names humanoid and four-legged robots as directly supporting the US military. Its reasoning is that the country uses more of this equipment than it makes and cannot meet rising demand. It sets no rate, no quota and no product codes, and names no country. Commerce and the US Trade Representative have 180 days to report on negotiations, and tariffs stay available after that.

Section 232 lets the Commerce Department decide whether a category of imports threatens national security. The president then picks the remedy as a separate step, which is why this one arrived with nothing attached.

Law firms including Mayer Brown and Thompson Hine pointed out that this is the third Section 232 case in three months where the administration chose talks over an immediate tariff. Brendan Schulman, Boston Dynamics' policy chief, expects about six months of them.

A3's numbers came out the same day. Companies in North America ordered 10,989 robots in the third quarter, worth $723.5M, up 24.8% in units. The growth came from the suppliers rather than the car plants.

A3 does not say where the robots were built, so the two documents cannot be lined up exactly. The order book does confirm the demand the proclamation describes, and right now imports fill much of it.

Tesla, Agility and GXO

Tesla told shareholders on 21 October that it has started building Optimus at Fremont, on a line where the Model S and Model X used to be made. Elon Musk called it early production with a gradual and uneven ramp, and set a target of 20,000 robots a week by the end of 2027. Tesla gave no unit count, no customer and no description of what the first robots will do. Quarterly profit fell 43%.

Agility Robotics' own filing projects about 800 Digit robots for all of 2027, so Tesla's weekly target is twenty-five times Agility's plan for the year. Agility also said it now expects about $465M from its SPAC merger, which is expected to close on 4 November, down from about $620M when the deal was announced in June. Foxconn is leading the private placement, which matters because Foxconn could also build the robots.

On 26 October Agility and GXO said Digit will go into two more GXO warehouses from early 2027. The first site, in Flowery Branch, Georgia, has moved more than 100,000 totes since June 2024. On GXO's August earnings call, chief executive Patrick Kelleher said "We have not achieved ROI on humanoids yet" and that the company is "a couple of years away from that". It is the clearest public statement on humanoid economics so far, and it comes from the buyer. GXO is expanding the pilot anyway, so it expects the robots to pay eventually, but it has not shown that they pay today.

AgiBot's Worken

At its Innovation Day in Shanghai on 29 October, AgiBot proposed a unit for robot work called Worken. It multiplies the number of robots doing real tasks, the hours they work without a person stepping in, and an efficiency ratio against a reference worker doing the same job.

Counting robots that work instead of robots that were built is the right idea. Every humanoid maker quotes shipments, and AgiBot led global humanoid shipments in the first half of the year. The weakness is that AgiBot defines every term, including the human reference. No other company has adopted Worken, and nobody has published a Worken figure that someone outside AgiBot could check.

Frontier models driving a real robot arm

Two writers on LessWrong, chrisleet and bira, tested nine frontier models on one real two-armed I2RT YAM robot from July to October. None of the models had robot training. Each one saw camera frames and wrote the commands that moved the arms, across six household tasks such as stacking bowls and capping a pen, with 20 trials per task.

They fitted a doubling time of 53.9 days, with a confidence interval of 36 to 59 days. The best models still score 45 to 48%, and extending the line reaches a human teleoperator around the middle of 2027.

A human teleoperator finishes a whole task in 25 to 40 seconds. The authors say the models rarely get confused about what to do. They miss grasps and misjudge distances. That matches my experience with perception on vehicles. Deciding what should happen is usually the easier half, and errors of a few centimetres against real geometry are where systems fail. Better reasoning in the language model does not fix that.

There are limits on how far to trust the slope. It is three months of data from one group, on one arm, through one software harness, and it is a blog post rather than a reviewed paper. GPT-6 Astra graded the trials and was also one of the models being tested, which the authors say openly. A second lab repeating it on a different arm with human grading would make the 2027 date worth taking seriously.

Why harnesses matter

A paper posted on 27 October, Your Policy Is Secretly a Harness by Jian Mu, Ruibo Shen and Yao Shu, shows why any leaderboard like this needs care. The harness is the software that feeds a model its observations and carries out its actions. The authors split an agent's score into the best result the harness makes possible, minus what the model loses through its own mistakes. Training only reduces the second part. Changing the harness changes both.

Their sharpest result is that reversing the order of the fields an agent reads, which changes nothing about the right answer, dropped a trained agent from 85.0% to 19.2% on a web benchmark. The work is on web and coding agents, and nobody has run the same test on a robot.

Robotics in Anthropic's S-1

Anthropic filed its public prospectus on 26 October, reporting $16.3B of revenue in the first half of 2026. Robotics appears three times, all in the summary and growth strategy. The one concrete sentence says the company is "extending Claude into robotics", through its Model Hardware Standard, which lets AI agents run lab and manufacturing equipment such as microscopes, liquid handlers and off-the-shelf robot arms.

The words "humanoid" and "embodied" appear in no Anthropic filing, and robots are not among the risk factors. A company lists what could go wrong in its risk factors and what it plans to sell in its growth strategy. Anthropic put robotics only in the second, and scoped it to lab and factory instruments rather than mobile robots.

Genie 4 and WorldOracle

Google DeepMind released Genie 4 on 27 October. It generates an interactive world in real time that you can change with plain-language prompts as it runs, and it is the first of its kind to generate sound with the picture. In one demo a Universal Robots UR5 holds a pipette and fills a lab plate. In another, Gemini 4 Argon drives a simulated Apptronik Apollo humanoid through a game of checkers. Google's own limitations section says visual detail and object permanence drift in longer sessions. Only trusted testers have access, so nobody outside Google can check the quality claims.

WorldOracle, from Princeton, UT Austin, the University of Toronto, NVIDIA and AgiBot, measures that kind of drift against real robot runs. It gives seven world models the first camera frame and the actions a real robot took, then compares the predicted video with what happened. Failed runs are included, so a model scores for predicting a dropped object correctly.

The models predict the robot's own arm two to three times better than the object it holds. My guess is that this comes from the data. The arm's movement is already in the actions the model is given, while the object's movement depends on contact physics the training signal barely covers. If that is right, the fix is supervision on contact and object state, and more video will not help much. The benchmark does not settle it. The write-up with these numbers is also labelled as sponsored by the WorldOracle team, and nobody has reproduced it yet.

JAM and robot policies that stop generating video

A world action model is a robot policy that predicts a short video of what its camera should see next along with the actions to get there. Learning to predict pixels teaches the model how things move, but generating frames costs time the control loop does not have. Three papers this week keep the training benefit and drop the video generation when the robot runs.

JAM, short for Just Action Models, by Yu, Huang, Li and Yamasaki, is the most interesting of the three because it proposes a different kind of model rather than a new module. The authors argue that the long debate between vision-language-action models and world action models mixed up three separate choices: how the backbone is initialised, what training objective is applied to future observations, and whether the machinery for predicting the future is still there when the robot runs. Only the last one made much difference in their experiments. JAM pairs a pretrained visual backbone with an action expert and adds one extra loss during training. The features the model computes from the current image have to predict frozen encodings of the real future frame, through a small head that is thrown away after training. No pixels are generated at any point. Against a matched FastWAM baseline, JAM scores 64.5% to 55.3% on LIBERO-Plus, a version of the LIBERO benchmark with changed lighting, cameras and objects, and the authors estimate 73% fewer training FLOPs per sample. Anyone who has shipped a model onto a vehicle will recognise this as distillation. You pay for the expensive signal once in training and ship the model that does not need it.

PDRVLA learns physical tokens by predicting future RGB, depth and optical-flow features, supervised by a frozen video world model, with no video at run time. RynnVLA-Latent learns latent actions from 30,100 hours of human, robot and simulation video that have no robot commands attached, and separates the camera's own movement from the movement in the scene.

This table is not a ranking. JAM's number comes from a matched comparison with FastWAM, PDRVLA uses its own data pipeline, and RynnVLA-Latent reports its result without tuning for the benchmark. All three are simulation only.

Rollback, from a sixteen-author team, keeps the world model at run time and uses it to rewind. When an attempt fails, it goes back to the decisions that mattered, tries other actions inside the world model, and compares the outcomes, which tells it which decision caused the failure. Real attempts on the robot keep updating the world model. Starting from teleoperated demonstrations, its policy passed average human success at ball tossing and tabletop curling within eight hours of training. Both tasks are short and have a clear outcome, so it is still open how this works when the deciding moment is harder to find.

The Tallinn crossing

At about six in the evening on 26 October, a Bolt delivery robot running on Starship Technologies' system entered a crossing on a red light in central Tallinn. A bus braked hard, standing passengers fell, and the city's transport company says up to nine people were hurt. The robot hit nobody.

Starship's head of operations, Pepe Aaviksoo, said a lit advertisement sat in the same line of sight as the traffic light, it was dark, and the robot's angle of approach made it worse. Andrei Novikov of Tallinn Transport said "The robot is at fault in any case."

This is a known failure for traffic-light detection. At night a bright sign can outshine the signal, and the check that would rule it out, that the light is at the wrong height and distance to be a traffic signal, is weakest when the robot approaches at an angle. The usual protections are a map of where the signal should be, a second sensor, and a rule to wait at the kerb when the robot is not sure of the light. Starship has not said which of these the robot had.

Boston Dynamics and the ICE contract

About 300 Boston Dynamics employees, nearly half of them signing as "Anonymous", asked leadership to cancel a $1.3M contract for four modified Spot robots for ICE. The company is keeping the contract. It says the robots go to Homeland Security Investigations, and that its terms of sale ban using Spot to harm or intimidate people or as a weapon.

That clause is the whole safeguard. It comes with no published audit right, no telemetry requirement and no stated penalty, and I do not know of any robot maker that has enforced a restriction like this against a customer.

Also this week

• Waymo chose the Toyota bZ4X as its next robotaxi, for Japan and the UK, and still targets a million paid rides a week by the end of 2026. Outside estimates put it near half that in early October.
• Bedrock Robotics raised a $600M Series C at a $4.1B valuation for autonomous excavators. Its CEO, Boris Sofman, said it is too early to give numbers on how many are deployed.
• Unitree's first report as a listed company shows third-quarter revenue up 70%, net profit down 17%, and research spending more than tripled.
• SK hynix expects the memory shortage to last until 2030 and named robots and self-driving cars as future demand. Tesla has already cut its AI5 chip from 144GB to 72GB of memory to reach Optimus volumes.
• Robot graders. RoboJev-27B, an open model that judges whether a robot succeeded, scored 0.890 AUROC on labs and robots it had not seen, below the 0.900 of the base model it was fine-tuned from. RoboSeed built a framework for ranking robot models and then declined to publish a ranking, because the available traces and protocols could not support one.
• Google put Gemini Robotics-ER 2 into public preview on Google Cloud.

What we're watching

• Does the 53.9-day doubling time hold up in a second lab, on a different arm and harness, with people grading the trials?
• Does anyone reorder the inputs to a robot policy, the way the harness paper did for web agents, and publish what happens to its score?
• Does Tesla publish an Optimus unit count or name a customer before the end-2027 target gets closer?