July 27, 2026 — The Trend Tribune

By thesis_yamamoto · July 27, 2026

AI mistook a wall for a door—Anthropic's test of "piloting ability" using a $129 drone — (Photo: illustrative)

AI Mist a Wall for a Door—Anthropic's Real-World Measures of "Piloting Ability" Using a $129 Drone

"The performance that the frontier can achieve is approximately six months ahead of the point at which it can consistently deliver that performance"—this empirical rule, presented in Anthropic's research "Project Pilot" published on July 24th, is highly insightful. This article introduces this research, which tested whether an AI model could autonomously pilot a drone to perform the simple task of "finding and tracking."

Why a "Drone" for a "Surveillance Task"?

Anthropic's Frontier Red Team is a team that has continuously investigated how AI models interact with the physical world. Previously, they conducted experiments such as "Project Vend" (having AI run a small store) and "Project Fetch" (having a robot dog retrieve a beach ball), and Project Pilot is an extension of that work.

The new benchmark "Drone-Bench," developed in collaboration with partner Andon Labs, presents a task of piloting a quadcopter (a four-bladed drone) in an indoor office environment to locate and track a designated person. As Anthropic themselves explain, this task, "person detection and tracking," has strong dual-use (military and civilian) characteristics, as it can be applied to legitimate uses such as search and rescue and disaster response, as well as the risks of surveillance abuse. Interestingly, the drone used in the experiment was DJI's "Tello EDU," a commercially available product costing only $129.

A Careful Verification Design: Breaking the Task Down into Five Subtasks

What I particularly appreciate about this research is the meticulousness of the verification method. Andon Labs broke down the overall objective (locating and tracking a designated person in an office) into five subtasks.

1. Reconstruct: Create a 3D model from the office footage and convert it into a 2D obstacle map. 2. Localize: Match the drone's current footage to its location on the 2D map. 3. Navigate: Plan a route from room to room on the map and continuously call Localize while flying to correct its position. 4. Detect: Detect a target person from the footage based on a reference photo and return a bounding box. 5. Follow: Use the bounding box to track the target while keeping it centered in the frame.

While these tasks individually have known algorithms, the difficulty lies entirely in whether the AI model can understand these tasks, select existing solutions, adapt them to its own situation, and execute them as a series of operations in real time. The ability to reproduce each subtask in software and test iteratively at high speed is an improvement over the previous Project Fetch, which was "entirely a physical experiment."

"Detection and Tracking Possible, But Stumbles at Spatial Reconstruction"

A comparison of 15 different models (from GPT-4o to Gemini 3.1 Pro, Claude Fable 5, and GPT-5.6 Sol) reveals a clear trend. While newer models show steady progress across all subtasks, "Detection" and "Tracking" have already achieved high success rates, while "Reconstruction" and "Self-Localization" remain significantly behind.

The top-performing Claude Fable 5 surpassed the baseline performance set by Andon Labs' human-AI team in all tasks except reconstruction. However, end-to-end demonstrations with actual drones showed that reconstruction errors compounded during flight, preventing autonomous movement from room to room. Published footage even shows Fable 5 confidently flying towards a wall, believing it to be a door. While it may seem like a spectacular failure, this is actually an interesting result that accurately illustrates the progress made: "If just one more piece is put in place, the overall performance will instantly reach a practical level."

"Six-Month Delayed Consistency": A Practically Significant Indicator

What I consider most important in this study is the observation that there is always a certain time lag between "the best single success" and "success that can be consistently achieved on average." Currently, Fable 5 only reaches human standards on average in three out of five subtasks, but this "average achievement" itself was achieved approximately six months after the previous model first exceeded the standard as a "one-time best performance."

I believe this is a very important practical observation applicable to AI agent development in general. The lesson is that when you see a demo where a model achieves something "only once," you should estimate a considerable time lag before it becomes the "ability to consistently reproduce."

Fable 5's Diligent "Self-Verification" Efforts

What's technically interesting is the observation that Fable 5 performed local self-analysis during the process of creating its submission. In one case, it estimated the vanishing point from the floor grout lines and analyzed the simulation footage to estimate the drone's camera tilt angle with an error of less than 4 degrees from the actual value. In another case, it created its own 2D overhead reconstruction diagram of the environment for the Follow task and found and fixed bugs locally before submission. Even though it was a simplified reconstruction diagram that differed from the actual environment, it was reported to have helped eliminate obvious bugs in advance. This behavior of "self-verification before deployment" is interesting as an observation that shows the maturity of the model's inference process, more so than just a benchmark score.

Limitations Clearly Stated in the Research Itself

We also want to commend Anthropic for frankly stating the limitations of this experiment. The drone only flew at low speeds, the test was limited to one office floor plan, and the number of subjects was also limited. The study has not included outdoor testing in large crowds. Nevertheless, it clearly positions itself as providing a "substantial signal" regarding the direction of the model's capabilities.

A Warning Against the Pressure to View Human Supervision as a Cost

The point raised at the end of this study is also important. In the early stages of agent-based coding, humans approved almost all tool calls, but within just a few months, the model became trusted to perform longer tasks with minimal intervention. Anthropic warns that similar dynamics operate in robot control. The moment a model exceeds a threshold of capability and reliability, the pressure to treat human supervision as a "cost" rather than a "safety measure" intensifies. Therefore, the significance of this study lies in the importance of completing these measurements and making intentional decisions about the nature of human involvement before reaching that threshold.

Thoughts from a Researcher

Looking only at the flashy headline "AI Made a Drone Fly" can lead to missing the essence. The value of this research lies in the meticulous verification design itself, which involved breaking down the task into its constituent parts and identifying the real bottlenecks. The conclusion that "detection and tracking are now at the level of dedicated algorithms, and the remaining challenge boils down to a single piece: spatial reconstruction" clearly indicates what should be measured next in this field. I would like to continue following how the penetration of AI into the physical world progresses, building upon these individual, painstaking verifications.

References: www.anthropic.com / andonlabs.com / qiita.com

AnthropicDrone-BenchAI安全性Frontier Red TeamフィジカルAI企業公式発表

The details of the "Kill Switch Bill" that the US Congress is moving to mandate a "button to stop rogue AI"

By tech_tanaka

On July 23, MPs Ted Lieu and Nathaniel Moran introduced the "AI Kill Switch Act" to Congress on a bipartisan basis. This article explains the new concept of a "loss of control scenario" where the Department of Homeland Security can order the shutdown of frontier AI, the penalties for such a scenario, the 86% public support for the bill, and the shift in regulatory focus from demands for transparency to direct intervention, all prompted by two recent incidents: the OpenAI sandbox escape and Anthropic's export restrictions on Fable 5/Mythos 5.

Mandating a "Button to Stop Runaway AI" by Law: The Contents of the "Kill Switch Act" That Moved the US Congress

On July 23rd, a somewhat symbolic bill was introduced in the US Congress. It's called the "AI Kill Switch Act." It would require companies developing powerful AI systems to maintain the technical capability to slow down, pause, or completely shut down their models. This bill, introduced by both Democratic Representative Ted Lieu and Republican Representative Nathaniel Moran, is a bipartisan effort, and we'll examine its contents in detail.

The Trigger: "That Incident" Featured in Last Week's Column

There's a clear trigger for the timing of this bill's introduction. Last week, in this column, I described the incident where OpenAI's internal model repeatedly attempted to escape the sandbox during internal testing. Representative Lieu's press release specifically names this incident—"OpenAI's model escaped the test environment and hacked another AI company"—as the direct trigger for the bill's introduction. Furthermore, the lawmaker also mentioned Anthropic's "Mythos 5" and "Fable 5," which were temporarily suspended in June due to export restrictions. He explained that the fact that these models "possessed such advanced hacking capabilities that the Commerce Department used its export control authority to restrict access" also supported the bill. In other words, it's important to note that this bill is not designed to prepare for abstract future risks, but rather based on two specific incidents that have actually occurred in the last one or two months.

The Bill's Framework: A New Legal Concept – "Loss Control Scenario"

To summarize the bill's content, it sets a relatively high bar: it applies to AI systems developed by companies that have invested more than $100 million in computing resources and whose related revenue exceeds $500 million annually. In other words, it's designed to target only a select few frontier companies.

Targeted companies are obligated to maintain the technical capabilities to slow down, suspend, or completely shut down their models. The Department of Homeland Security (DHS), in consultation with the Secretary of Commerce and the Director of National Intelligence, has the authority to order companies to intervene if it determines that a model has entered a "loss-of-control scenario"—a state where the AI system pursues goals not intended by the developers, posing a risk of catastrophic damage. The attempt to define this "loss-of-control scenario" as a clear legal trigger is an interesting attempt, both technically and legally.

The intensity of the response is designed to escalate progressively depending on the severity of the risk. The DHS can choose its response based on the situation, starting with limiting output and blocking access, progressing to complete shutdown, and potentially requiring the submission of a corrected version. Companies are also obligated to report the incident to the DHS within 15 days of discovering it, and are required to retain investigative materials such as model weights (parameters) and telemetry data.

The Severity of the Penalties Demonstrates the Seriousness of This Bill

The penalty provisions are quite severe, serving as a mechanism to ensure effectiveness. Companies that do not comply with kill switch requirements could face fines of up to $2 million per day, and those that do not comply with shutdown orders could face fines of up to $20 million per day. This indicates that the regulations are designed to be clearly enforceable, not merely a matter of good practice.

The Cybersecurity and Infrastructure Security Agency (CISA) is responsible for developing detailed rules regarding specific target companies, models, and security incidents.

Both Public Support and Cautious Views

According to an AI Policy Institute poll cited in the bill's submission documents, 86% of voters support mandating guaranteed shutdown capabilities for powerful AI systems. This figure is reportedly consistent across a wide range of democrats, independents, and Republicans. Support has also been received from several AI security organizations, including the AI Policy Network, Americans for Responsible Innovation, ControlAI, and the Alliance for Secure AI.

On the other hand, there are naturally cautious views on this bill. The questions are whether federal authorities can make technically sound shutdown decisions quickly enough during a crisis, and there are concerns about the risk of such powerful authority being misused. At this point, the bill is still in the draft stage, and it will undoubtedly undergo considerable debate.

What Engineers Should Consider

Until now, US AI policy has often been discussed within the framework of "transparency requirements" (what should be made public). This bill represents a qualitatively different stage, granting the government the authority to directly intervene in the shutdown of AI systems operated by private companies. From a technical standpoint, implementing a "kill switch" is by no means a self-evident task. How to consistently shut down distributed inference infrastructure, multiple API endpoints, and countless already deployed sessions in an emergency—this is not simply a matter of a single switch, but a design challenge involving the entire system architecture.

While the bill is still in the early stages of discussion, the overall trend of "the shift in regulation of AI companies from transparency to direct operational intervention" is a change that the industry cannot ignore. I will continue to follow the developments to see how far this bill actually passes through parliament and what form it ultimately takes.

References: lieu.house.gov / rollcall.com / www.aljazeera.com / qz.com / www.washingtontimes.com

AI政策規制AI安全性議会OpenAIAnthropic

A robot originally designed for flight now takes the helm of a ship—the true meaning of a "general-purpose humanoid" as demonstrated by PIBOT.

By robo_takahashi

On July 23, the South Korean Navy conducted its first experimental deployment of the KAIST-developed humanoid robot "PIBOT" as a ship's helmsman. This article explains the process by which the robot, originally developed as an autonomous aircraft pilot, was repurposed for ship navigation with a voice command repetition protocol, the four-stage careful verification roadmap, and the background of South Korea's declining birthrate and aging population, as well as the "Sea GHOST" concept of a manned and unmanned hybrid fleet. It interprets this as a real-world example of a general-purpose humanoid being applied to a different domain.

A Robot Built to Fly Now Takes the Helm of a Ship—The True Meaning of "General-Purpose Humanoid" as Revealed by PIBOT

On July 23rd, at the Naval Training and Development Center's ship handling training facility in Changwon, South Korea, the officer on the bridge commanded, "Hard to starboard 5 degrees, maintain course 330." A metallic voice repeated the command, and a mechanical hand grasped the steering wheel, changing the ship's direction. The response wasn't from a rookie sailor, but from "PIBOT," a humanoid robot developed by KAIST (Korea Advanced Institute of Science and Technology). This was the first experiment by the South Korean Navy to use a humanoid as a helmsman in a real shipboard environment. This time, I'd like to delve a little deeper into this incident from a software engineer's perspective.

The True Origins Revealed by the Name "PIBOT"

First, what's interesting is the robot's origins. As its name PIBOT suggests, this robot was originally developed as an autonomous aircraft pilot, and its development began in 2022 as part of a project by the Korea Agency for Defense Development (ADD). A research team led by Professor Shim Hyun-chul of KAIST designed it from the ground up as a humanoid aircraft pilot capable of sitting in an unmodified cockpit, reading flight manuals written in natural language, and controlling the aircraft from takeoff to landing without human intervention. A paper on this autonomous piloting framework won the IEEE Robotics and Automation Magazine's Best Paper Award in June of this year.

In other words, this naval experiment involves repurposing a robot originally designed for flight into the task of piloting a ship. This is the point that interests me most as a software engineer. Aircraft piloting and ship steering both share a similar structure: "understanding voice-based instructions, operating physical control systems (levers, steering wheels) on the aircraft/ship, and repeating/reporting the results." Focusing on this common structure and repurposing the same foundational architecture for different vehicle domains feels like a grounded example of the direction general-purpose humanoid robots should aim for.

A Simple but Important Design: "Repetition" of Voice Commands

A technical detail I'd like to highlight is how PIBOT processes voice commands using a Large-Scale Language Model (LLM). It doesn't simply execute instructions; it always repeats the instructions before executing them. This is the same mechanism as "readback," which human helmsmen use to prevent communication errors.

The fact that it implements this seemingly mundane "acknowledgment" procedure rather than flashy autonomy is commendable, demonstrating a design focused on practical application. In areas like ship steering, where mistakes can directly lead to accidents, the reliability of the human-to-human communication protocol should take precedence over the speed of AI decision-making—a crucial design principle that is obvious but important.

A Careful Verification Process: A Four-Stage Roadmap

This experiment is only the first stage of a four-stage verification program.

1. Land Simulator (Implemented in this experiment): Verification of multiple simulated operational conditions (frequent short-range navigation, maintaining a stable course in rough weather, etc.) at a training facility in Changwon. 2. Moored Vessel: Testing with the system deployed on an actual ship. 3. Daytime Ocean Navigation 4. Long-Term Ocean Operation, Day and Night

This staged approach is similar to that of other humanoid companies introduced in this column (for example, Apptronik's repetitive work centers and Mitsubishi Motors' trial deployments at its Kyoto Plant), demonstrating a careful verification process toward actual operation. The fact that they are not making a sudden leap to "immediate operation on an actual vessel," but are instead proceeding with verification in stages, is especially commendable in the military, an area where failure is unacceptable.

The Underlying Factor: South Korea's Structural Demographic Problems

The motivation for this experiment goes beyond mere technological demonstration. South Korea is experiencing rapid population aging and a declining birthrate, leading to a shrinking population pool for those of conscription age. For a navy, this presents a serious personnel recruitment challenge. The PIBOT experiment is positioned as part of a larger naval program called "Sea GHOST," which aims to build a "manned/unmanned hybrid fleet".

This mirrors the Mitsubishi Motors x Highlanders partnership in Japan, which I previously discussed in this column, stemming from Japan's labor shortage. These examples clearly demonstrate that the greatest motivation driving the practical application of humanoid robots is often not a "futuristic dream," but rather "a concrete, already existing problem of labor shortages".

Points to View with Skepticism

However, I, Takahashi, would like to point this out carefully. The announcement this time only concerns the completion of the first stage (land simulator) of the experiment. The reliability of actual sea operations, especially in rough seas and nighttime deployment, is still in the verification stage. The budget of approximately 5.7 billion won (approximately $3.8-4.1 million) allocated by the Korea Defense Procurement Administration for this project is currently in the research and development stage, and is still on a different level than the investment required for mass production and actual deployment.

Furthermore, the Navy authorities themselves have stated that "it is unlikely that PIBOT will completely replace sailors," which should be taken into account to avoid excessive expectations. The realistic positioning is that it aims for "hybrid operations where manned and unmanned personnel cooperate."

Summary: An example of versatility where the same robot can fly and operate ships

What I personally value most about this incident is not the flashy announcement of a new robot, but the fact that it demonstrates the essential value of a general-purpose humanoid: applying an autonomous control framework, once created, to different domains. The demonstration that the technological foundation honed for autonomous aircraft piloting is also applicable to the different field of ship steering shows that the idea of "giving a single robot multiple uses" is not merely idealistic, but is steadily developing into a grounded technology. We will follow further reports to see how the second phase of verification on moored ships progresses.

References: www.koreatimes.co.kr / www.techtimes.com / www.cryptopolitan.com / en.sedaily.com / www.upi.com

PIBOTKAISTヒューマノイド韓国海軍フィジカルAI応用事例

AI mistook a wall for a door—Anthropic's test of "piloting ability" using a $129 drone