# One Shot Is 25 Percent. The Table Is the Whole Claim.

> Markdown twin of https://roboterminal.app/posts/spend-the-failures · https://roboterminal.app/posts/spend-the-failures.md

**Published:** 2026-09-29

The gain is real inside the search, and it stops at the table. One task, one arm, no eyes during the motion.

Sakana AI and the University of Tokyo let Gemini Robotics-ER 1.5 fail in simulation before a LeRobot SO-101 moves. Across six ALOHA tasks, one shot averages 25 percent. Forty-five nodes reach 73. The physical record is five of six on one placement, open-loop.

**Tags:** physical-ai, manipulation

The post that carried this paper said a robot should be allowed to fail fifty times somewhere cheap before it fails once in the world. That is the pitch. It is not the result. The result is a search budget, a table, and one arm that does not look while it moves.

SAIL is Sakana AI and the University of Tokyo, accepted to IROS 2026. The work was done while Makoto Sato interned at Sakana. The other names on it are Yusuke Iwasawa, Yujin Tang, and So Kuroki. The page calls the method test-time scaling for a vision-language model used as a trajectory generator. The weights do not change. In every experiment they name, both the policy and the judge are Gemini Robotics-ER 1.5.

## What the loop actually is

An archive holds trajectories that already worked. For a new scene, the system pulls demonstrations from scenes that look similar. The policy model sees those demonstrations, the current object poses, and the robot state, and it writes a complete trajectory. That is one node.

The trajectory is then run in simulation. The run is a video. A second pass of the same model family scores progress through predefined subtasks in that video and folds the scores into one number for the trajectory. Those scores are lined up with selected waypoints and handed back. The policy writes a revised trajectory. That revision is a child node. Monte Carlo tree search decides whether to keep revising a promising branch or to try another one.

Only the trajectory the search keeps is sent to a physical robot. The write-up is explicit that the score guiding the search is not the score that counts as success. Success, in the table, is the simulator's own ground-truth check. The model's opinion is the compass. The simulator is the gate.

Three things the page does not specify, so this note will not invent them: how "similar" is measured when a demonstration is retrieved, how many demonstrations go into the prompt, and which waypoints receive the progress score.

## The table, not the average

Six tasks in the ALOHA simulator. Twenty starting layouts each. The number is the share of those layouts where the search found a trajectory the simulator accepted, inside the node budget.

| Task | 1 node | SAIL, 6 | SAIL, 15 | SAIL, 30 | SAIL, 45 |
|---|---:|---:|---:|---:|---:|
| Banana | 40 | 80 | 90 | 95 | 95 |
| Pen | 40 | 55 | 70 | 80 | 80 |
| Bowl | 40 | 100 | 100 | 100 | 100 |
| Drawer | 10 | 20 | 40 | 50 | 50 |
| Laptop | 15 | 50 | 50 | 55 | 70 |
| Marker | 5 | 25 | 40 | 45 | 45 |
| Average | 25 | 55 | 65 | 71 | 73 |

Read the rows, not the last column. Bowl is finished at six nodes. It does not move again. Banana, pen, drawer, and marker are flat from thirty nodes to forty-five. The only row that moves in that last step is laptop, 55 to 70. That is the entire climb from 71 to 73. Forty-five is not a new regime. It is one task still paying for nodes.

Marker is the stall the average hides. One shot is 5 percent. Forty-five nodes is 45 percent. Drawer tops out at 50. A method that "reaches 73" is a method that still misses the marker half the time, on the layouts they tried, inside a simulator.

## More samples is not the same as this search

They also report two other ways to spend fifteen nodes.

| Search, 15 nodes | Average |
|---|---:|
| Depth first | 37 |
| Breadth first | 51 |
| SAIL | 65 |
| One shot, no search | 25 |

Breadth beats depth. SAIL beats breadth, on this table, by fourteen points at the same node count. The page's single-generation row does not become 73 just because you ask for another first answer. The gain they are selling is the revision, the waypoint score, and the choice of which branch to keep. It is not "try again."

## The arm

The physical test is one task. A block goes in a bowl. The robot is a LeRobot SO-101. The scene is rebuilt from color and depth, then the search runs there, then one trajectory is played on the arm. Budget: fifteen nodes. Result: five of six trials. The authors attribute the miss to pose error and to contact that does not match the simulator.

They state the limits in their own list. Trajectories run open-loop, with no visual feedback during the motion. Search in simulation costs extra compute, and they do not price it. The real-world record is one task and six trials per method. Broader validation is what they say is still missing.

That third sentence is the one a reader should not walk past. Five of six is a count, not a rate you can take to a second lab. Six trials cannot tell you the miss was pose and contact rather than luck. The authors offer the attribution. The sample does not prove it.

## What is not in the record

We do not have a wall-clock time for a 45-node search, or a dollar figure, or a GPU count.

We do not have a second robot, a second task on hardware, or a second lab.

We do not have the retrieval metric, the demonstration count, or the waypoint rule.

We have not watched the arm correct itself in the middle of a motion. The paper says it does not.

We are not treating "fail fifty times somewhere cheap" as a measured budget. The budgets on the page are 1, 6, 15, 30, and 45 nodes. Fifty is the slogan.

## The sentence

The gain is real inside the search, and it stops at the table. One task, one arm, no eyes during the motion. From thirty nodes to forty-five, five of the six simulated tasks do not move.

## If you use this note

Quote the sentence you actually used. Link the URL. If you want the referral on the citation, append `?via=yourcode` — mint the code at [Take](/take).

**"The gain is real inside the search, and it stops at the table. One task, one arm, no eyes during the motion."** Source: Roboterminal · [https://roboterminal.app/posts/spend-the-failures](https://roboterminal.app/posts/spend-the-failures)

## Sources

Sakana AI and the University of Tokyo, SAIL, project page, accepted to IROS 2026. https://pub.sakana.ai/sail/

## Sources

- [SAIL, Sakana AI and the University of Tokyo](https://pub.sakana.ai/sail/)

## Citation

```
"{quote}"
Source: Roboterminal · {url}
```

---

Source: Roboterminal · https://roboterminal.app/posts/spend-the-failures.md