Series 6 ended with v6 at 22%. Most of its failures looked the same: it reached the tube, stopped with the jaws open and froze. This article compares three versions built on that model, a, b and c, and the one change behind each.
How do v7a, v7b and v7c compare?
Share of 50 test runs where the model picked up the tube with no help. The 4 second rule hands over to the rescue after 4 seconds without movement. The relaxed rule waits 10 seconds and counts gripper movement as progress. v7c's guard holds the arm still on purpose, so only the relaxed rule applies to it.
| What changed | Retrained? | Main failure | |
|---|---|---|---|
| v7a | Added v6's 50 runs as a fourth round of rescues | Yes | Froze above the tube (34 of 42 runs) |
| v7b | Relabelled every rescue the way DAgger intends | Yes | Closed and lifted, but the tube slipped out (27 of 50) |
| v7c | Added a grip guard between the model and the arm | No | Still empty grips, fewer of them (16 of 50) |
How has each model done so far?
Share of test runs where the model picked up the tube with no help, on the 4 second rule. ACT is from Series 2, SmolVLA v1 to v3 from Series 3.
Why did v7a get worse?
In every round since Series 4, a rescued run went into training as only the rescue skill's frames. The model's frames, including the seconds it sat frozen above the tube, were dropped. The rescue itself starts by rising and backing off. So the exact state where the model gets stuck never got a label, and the nearest examples said "back off". More rescues made it worse.
What did v7b change?
The red block is what DAgger is meant to label: the states the model actually reached. L is the rescue's lined-up pose just before its final descent.
The model's last 3 seconds before each takeover now carry the label "move to the lined-up pose", followed by the rescue's descent, close and lift. Same runs, same everything else. 94 of 107 rescues were relabelled. The freeze disappeared: 3 stalls in 50 runs, against 34 in 42 for v7a.
v7b's fastest run, 2× speed. It reaches, closes and lifts the tube in 6.6 seconds.
Why did v7b still miss?
It now committed to the grasp, but in 27 runs the tube slipped out as the arm rose. The cause was timing. After the jaws stop moving, a miss waited a median 0.28 seconds before lifting. Successes and the rescue skill waited about twice as long.
Pause between the jaws settling and the arm starting to rise. In 4 misses the arm rose before the jaws had even finished closing.
The data was not the cause, because every source in it waits about half a second. SmolVLA plans 50 steps, about 1.7 seconds, from one snapshot and plays them without looking again, so it lifts on schedule even when the close runs slow. The rescue checks the gripper every frame and waits.
What did v7c change?
v7c is v7b with its weights untouched and a small rule between the model and the arm. The model still decides where to go, when to close and how far. At the grasp the guard:
- caps how fast the gripper closes, at the rescue skill's pace;
- holds the arm still while the gripper is closing;
- lets the lift through once the gripper reading has been steady for about half a second.
| v7b | v7c | |
|---|---|---|
| Model alone, relaxed rule | 20 / 50 (40%) | 24 / 50 (48%) |
| Rescued | 27 | 24 |
| Failed | 3 | 2 |
| Median success time | 10.0 s | 14.0 s |
The guard was tuned during the test. On its own it did best, 13 of 22 (59%) at v7b's speed. A stricter version that slowed both the close and the lift got 6 of 20: it held the arm so long that the model's plan moved on, and the arm closed, reopened and retried until time ran out. Over 50 runs the gain from v7b to v7c is not yet significant (p = 0.55).
What comes next?
A hand-written rule helps, but it has to guess how long the model's plan will wait, and holding too long fights the plan. Next is a small learned correction network next to the frozen model, trained on the rescues: at each moment, what the rescue did minus what SmolVLA proposed. It sees the model's command every frame, so it can learn when to wait.