Teachers: the kicks

The first kicks, all from my motion capture at the gym on 25 October (Muay Thai 03, take_01, two phones). I picked them off the preview: the right low kick at 3 s, the left low kick at 11 s, the right mid kick at 22 s and the left mid kicks at 27 s, then the two high kicks at about 33 s and 40 s. I wanted each kick trained as its own technique, with no blending between them. The one exception is the left mid kick, which mixes my two left mid kicks of different heights by the height aimed at. Each is aimed a little higher, lower, left and right, but not nearer or farther: “it doesn’t need a huge range”. The left kicks are switch kicks: a hop switches the feet, the kick, and the switch back. The techniques are on the low kicks, mid kicks and high kicks pages.

The runs

run	from	what	outcome	machine-hours, cost
lowkickR1	distill4-131.1M	right low kick to the thigh	223.2M: air 6/30, bag 0/30; my review pending	2.6 h, $2.36
lowkickL1	distill4-131.1M	left (switch) low kick	224.6M: air 30/30, bag 0/30; my review pending	2.6 h, $2.35
midkickR1	distill4-131.1M	right mid kick to the body	223.5M: air 30/30, bag 0/30; my review pending	2.6 h, $2.35
midkickL1	distill4-131.1M	left (switch) mid kick, two heights	228.9M: air 29/30, bag 0/30 (stops 26–35 cm short); my review pending	2.6 h, $2.35
highkickR1	midkickR1-223.5M	right high kick to the head	209.7M: hits at the bag and in the air below the head; my review pending	2.6 h, $2.28
highkickL1	midkickL1-228.9M	left (switch) high kick to the head	199.9M: bag 5/6 at most aims, no air hits at head height; my review pending	2.6 h, $2.28
lowkickR2	lowkickR1-223.2M	the bag on the target, no side-bend after the kick	kept 286.1M: air 0/30, bag 12/30	2.6 h, $3.15
lowkickL2	lowkickL1-224.6M	the same	kept 227.4M: air 30/30, bag 30/30	2.6 h, $2.35
midkickR2	midkickR1-223.5M	the same	kept 221.9M: air 23/30, bag 27/30	2.6 h, $2.35
midkickL2	midkickL1-228.9M	the same	kept 275.4M: air 18/30, bag 5/30	2.6 h, $3.15
highkickR2	highkickR1-209.7M	the kick’s height before it lands	kept 273.2M: air 35/36, ankle at head height on the bag	2.6 h, $3.05
highkickL2	highkickL1-199.9M	the same	kept 271.2M: air 30/36, ankle at head height on the bag	2.6 h, $3.06

Round 1: the low and mid kicks

The four trained in parallel on RunPod RTX 4090s ($0.89/h; no 5090 was free) from 07:49–07:51 UTC on 26 October, to their 2.5-hour cap, about 223–229 M steps each. Each was warm-started from the student distill4-131.1M, with distill4 as the teacher for everything except the kick itself. distill4 had never kicked. They were launched from my Windows PC with a new mode that follows the pod only through S3, with no ssh. The first try of each died after about 8 minutes because the PC’s S3 downloads were refused. That was fixed and the four relaunched (about $0.08 each, in the costs). A kick is scored at the ankle, and the shin and foot are what land on the bag.

The gate ran the next morning on the PC, six seeds at the centre and the four ±10 cm aims, in the air and at the bag:

	air hits	air: off target, speed	bag	hips after
lowkickR1-223.2M	6/30	3–8 cm, 2.6–4.1 m/s (slow)	0/30, 5–11 cm off	turn up to 35°, back to ~2°
lowkickL1-224.6M	30/30	4–7 cm, 6.8–7.6 m/s	0/30, 2–6 cm off but braking into it	turn 72–82°, back to 2–4°
midkickR1-223.5M	30/30	4–6 cm, 5.9–7.5 m/s	0/30, 6.5–12 cm off	turn 43–56°, back to 3–8°
midkickL1-228.9M	29/30	1.5–8 cm, 4.1–5.3 m/s	0/30, 26–35 cm short	turn 42–65°, back to 1–5°

None of them fell, none took a step after the kick, and none fell in the balance test with presses and shoves. The right low kick reaches but doesn’t snap. Not one kick hit the bag.

The high kicks

On 27 October I gave the go for the high kicks, aimed at the opponent’s head (1.56 m) and at points to the sides and below it, but never above: “no need to kick above that”. The aims are down 15 cm and 10 cm to either side. The other glove is moved to the ear as the kick lands. In my capture the shin peaks at 1.42 m (right) and 1.47 m (left), so the hips are raised up to 16 cm to reach the head. They had their own env, built after the round-1 videos, with the bag on the target and the side-bend term from round 2 (below). They were warm-started from the mid kicks and trained on two RTX 4090s from 01:19 UTC on 27 October to the 2.5-hour cap.

At the end the final checkpoint and the training log came down from the pod as 0-byte files, so the last snapshot of each was used (209.7 M and 199.9 M). They are on the PC only, not on S3 yet. The gate:

• highkickR1-209.7M: air hits 5/6 at the head and 6/6 15 cm below it, 2–3/6 at the sides; at the bag 2/6 at the head, 6/6 on the low row, 1/6 at the sides; 4–12 cm off; no falls, no steps. The hips turn up to 127–139° in the air and end 16–21° off.
• highkickL1-199.9M: in the air 0/6 at the head and its sides, 4/6 15 cm below; at the bag 5/6 at every aim but the low right (2/6), 1–12 cm off; one seed in six falls in every row.

Round 2, prepared

I watched the round-1 videos and saw why the low and mid kicks miss: “the green dot is completely outside the bag so of course it’ll miss the bag … let’s put the green dot completely inside the bag at the appropriate height depending on the kick”. I also said: “after the kick the fighter loses balance, it bends its body left and right, that should not happen, in the reference it balanced well”. The chest had bent 14–35° the other way from my reference on the way back down. Two changes, for all four kicks:

• The bag on the target. The bag now hangs with its axis through the target, 17 cm inside its face, at the kick’s own impact height. Before, the target was at the front of the shin, outside the bag.
• No side-bend. A reward term (weight 2.0, an 8° Gaussian) for the chest and pelvis matching the reference’s side-bend from the hit to the end of the kick.

lowkickR2, lowkickL2, midkickR2 and midkickL2 are each warm from their round-1 kick, with the same settings otherwise. They trained on RunPod (two RTX 5090s at $1.19/h, two RTX 4090s at $0.89/h) to the 2.5-hour cap. The gate, on 28 October:

	air hits	bag hits	side-bend after the hit
lowkickR2-286.1M	0/30 (reaches the target but stops there, doesn’t kick through)	12/30	4–8°
lowkickL2-227.4M	30/30	30/30	4–6°
midkickR2-221.9M	23/30 (the hips over-rotate, 115–127°)	27/30	8–11°
midkickL2-275.4M	18/30	5/30 (only at the high aim, but it reaches the bag now)	4–6°

Every kick now reaches the bag and bends far less after it. None fell or stepped. I watched the four together and said: “Looks good, please mark it keep.”

The high kicks, round 2

On the round-1 videos I saw that “the kick seems to be hitting the bag before it gets the height”. It had three causes. The target sat off a roundhouse’s path. The bag paid for the push wherever the ankle met it, so on the bag rows the ankle topped out at 1.1–1.3 m for a 1.6 m target. And the bag was in the leg’s way. highkickR2 and highkickL2 score a hit along the ankle’s way in over the last 0.1 s, pay the bag’s reward only when the ankle is on target, and start with every kick in the air until 60 % of them hit. They were warm from round 1, with distill4 as the teacher, on two RunPod RTX 5090s ($1.19/h) for 2.5 hours.

• highkickR2-273.2M: air 35/36; at the bag the ankle tops at 1.53–1.59 m for 1.64 (round 1: 1.21–1.28).
• highkickL2-271.2M: air 30/36 (one seed in six falls, as in round 1); at the bag the ankle tops at 1.52–1.57 m for 1.65 (round 1: 1.26–1.30).

Neither counts a bag hit yet: the shin sweeps across the bag’s face and hardly moves it. But the height is fixed. I said: “looks good, consider it a keeper.” The high kicks went into the student in distill9. The low and mid kicks are kept teachers, not in a student yet.

Every run

From the run records in kakuto-ml: each run's machines, their billed hours and cost (EC2 runs are estimates from the run logs).

← All the work logs