Techniques · Evasions
The slip and uppercut
Slip to the left and come up with the lead uppercut.
Where it stands
| Technique | Teacher kept | In the student | Own runs |
|---|---|---|---|
| slip and uppercut | slipup1-204.5M24 October 2026 | not yet | 3 |
The slip and uppercut is from my motion capture, Muay Thai 02, take_01, 20.75–22.30 s: slip to the left, then the lead uppercut. It was one of four evasion teachers started within a minute of each other on 25 October, each cloned from the student distill4 (work log).
The first, slipup1-204.5M, scored 8 of 8 at the gate. On video it was less convincing. It played my clip 1.42 times as fast on average and threw the uppercut about 0.5 s early, so the slip before it was squeezed out. The cause was the punch clock: built for punches, it lets the policy play a reference up to twice as fast and pays for reaching the target sooner. Only the jab is held to its natural pace. The gate doesn’t check timing, so a rushed move still scores.
I asked for the slip to move the head further to the side. The reference was widened to hold the head 25 cm out, and slipup2-196.6M learnt it by bending its trunk sideways 52°. It looked more like a football headbutt than a slip, and it failed. A rebuilt version of the reference, with the trunk kept straight, looked physically wrong to me and was never trained. I went back to my original capture.
slipup3 started from slipup1 on that original capture, with the target moved down to the chin (1.49 m; it had been at eye height, 1.61 m). Its machine ran for 2.18 hours, but the record has no outcome and no model was kept from it. slipup1 stays the teacher until a better one exists, and the slip and uppercut is not in the student. Next: the punch clock held to my capture’s pace for evasions, and a gate that measures timing and shape, not only the hit.
What I think, as a coach
On slipup1: “the uppercut still looks good, so keep that one”, “but actually even that one looks like a big uppercut [more] than a slip”.
On slipup2: it “looks more like a football headbutt, it just stays there … It should recover straight away, like a U shape, but the head should be looking straight. The head doesn’t tilt, therefore the torso doesn’t tilt”. And “the target should be where the chin is, not higher than that.”
When the edited references looked wrong: “Keep to my original reference, like the one from motion capture.”
Every attempt
Every run that trained it on its own, in order.
| Run | Outcome | Machines | Hours | Cost |
|---|---|---|---|---|
slipup1-2026102524 October · slip and uppercut · kept slipup1-204.5M · Teachers: slips, rolls and the lean back | Kept | 1 | 2.2 | $1.48 |
slipup2-2026102525 October · slip and uppercut · Teachers: slips, rolls and the lean back | Failed | 1 | 2.1 | $1.41 |
slipup3-2026102525 October · slip and uppercut · Teachers: slips, rolls and the lean back | Not kept | 2 | 2.3 | $1.51 |
| 3 training runs | 4 | 6.6 | $4.39 |
How the models descend
Its own runs, each with an arrow from the run it started from; faded boxes are models from elsewhere it started from or learnt from. Tap a run for its details. The whole history is on the lineage page.