# Stanford VIA study reports a coding agent controlled a robot in 24 of 25 trials

> Commentator Nate B Jones describes a Stanford study in which a general coding agent guided a physical robot through five tasks without robot-specific training.

- Status: reported (One outlet, not independently verified yet.)
- Topic: Research
- Published: 2026-11-18T15:00:36+00:00
- Updated: 2026-11-19T17:09:15.801089+00:00
- Web page: https://ai-news.ikangai.com/story/stanford-via-study-reports-a-coding-agent-controlled-a-robot-in-24-of-25-trials/
- Seen in: Nate B Jones
- Change: research (Research)
- Applies to: VIA

## What changed

A Stanford study reports that a general-purpose coding agent called VIA controlled a physical robot through five tasks without robot-specific training, completing 24 of 25 attempts.

Stanford researchers used a general-purpose coding agent to direct a physical robot through five tasks. According to commentator Nate B Jones, who described the study, the agent completed 24 of 25 attempts with no additional robotics training. The system is called VIA, for Visual Interface Agent.

The mechanism splits the work. The agent is shown a camera view of the scene through a visual interface and places a virtual manipulator target where it judges the real gripper should go. It can adjust that target before anything moves. Once satisfied, it commands the robot's controller to move the physical arm to that point. The agent then checks the outcome, asking whether the gripper actually caught the handle and whether it needs to adjust before pulling. Jones stresses that each decision follows an observed result rather than a fixed sequence. His drawer-handle example, reach, grab and pull, illustrates the process and is not a reported trial.

The agent does not calculate each movement of the arm moment by moment. The controller moves the hand toward the chosen target, and selecting that target is a significant part of the job. Jones compares this to human reaching, where a person selects a target and the arm gets there without the person planning each muscle. He also notes that a correctly placed virtual gripper does not guarantee a good grip on a drawer, which is why the physical check after each move matters.

Jones calls the study small, but says he has seen similar work suggesting that an existing general-purpose agent can handle physical tasks without customization for a specific robot. He cites one example from the paper, a wiping task. The robot's view of the table was obscured, so the agent moved the gripper aside to check whether the surface was clean. In one attempt, holding the napkin was awkward, so the agent set it on the table and pushed it with a closed gripper.

Jones ties the result to Richard Sutton's Bitter Lesson, the idea that general methods that scale tend to outperform hand-built domain knowledge. In his reading, the agent's broad training probably already contained some robotics-relevant knowledge, and the researchers only had to connect its general abilities to a machine. He also places the study alongside OpenAI's reorganization of its robotics work. Jones said the company confirmed that its robotics researchers are moving to basic research while applied robotics and hardware move to a separate division, and that reports discuss financial pressures. He argues the technical logic is that improvements to a general model can carry over to robot control.

Jones's main caveat concerns speed rather than accuracy. He cites an October study, which he says is not the same setup as VIA, in which about 90% of the time on a G1 robot went to the model's thinking, and successful manipulation attempts took more than an hour. He also cites a November study called RUA, in which the base model stayed unchanged while the simulation score improved because of software built around it. His point is that surrounding infrastructure matters, and that a slow response can make a robot demand more human attention than it saves.

## Why it matters

If general agents can drive robots without task-specific training, companies building robots may get more capability from upgrading their underlying AI models than from custom robotics models. The study Jones cites is small. The result is not yet proof for commercial deployments.

---

AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it.

Cite as: IKANGAI AI News, "Stanford VIA study reports a coding agent controlled a robot in 24 of 25 trials", 2026-11-18, https://ai-news.ikangai.com/story/stanford-via-study-reports-a-coding-agent-controlled-a-robot-in-24-of-25-trials/
