Teachers: the uppercuts, second round

The first round of uppercuts over distance ended with my note on the video: “After the uppercut the fighter turns 90 degree to the right. It needs to keep facing the opponent.” The cause was in training: the home heading was re-set at every press and every shove recovery, and the opponent was moved round to wherever the fighter faced, so nothing paid for facing one way. These four runs are the fix, and then a second problem it uncovered. All four ran on RunPod in EUR-IS-1 on 26 October, each warm-started from an earlier uppercut, with distill4-131.1M as the teacher for everything except the uppercut itself. The technique is described on the uppercuts page.

The runs

run	from	change	outcome	machine-hours, cost
upL2 (lead)	upL1-261.3M	--face-fixed: the home heading set once per episode, the targets and the opponent kept in it, a facing term (×0.3)	failed. Square at the bag, but the far uppercut into the air still ends 8–45° off at 3 s; air hits 33–44/54, no better than upL1; then it staggered on video	3.1 h, $2.79
upR2 (rear)	upR1-209.7M	the same	kept 209.7M: air 54/54, ends 3–17° off square (upR1 up to 30°), no falls; bag hits 25/54 (upR1 38)	2.5 h, $2.70
upL3 (lead)	upL2-209.7M	facing term ×1.0 instead of ×0.3	not kept: the far air uppercut still ends 14–45° off; one video seed turns 99°	2.6 h, $2.16
upL4 (lead)	upL2-209.7M	feet kept down after the punch, half the presses into the air (was a quarter)	kept 157.3M: air 54/54 at 1.4–2.5 cm, 0 steps after the punch, ends 1–12° off in the air	2.9 h, $2.40

upL2 and upR2 trained side by side on two RTX 5090s ($1.19/h) for their 2-hour cap, about 214 M and 218 M steps. upL3 and upL4 ran on RTX 4090s ($0.89/h) because no 5090 was free. The costs include the gate pods and the video renders.

Facing: the weight was not the lever

The gate now measures the hips’ turn 1–3 s after each press. With --face-fixed, the rear uppercut came back square at every aim, and so did the lead uppercut at the bag (2–5° at 3 s for upL2-209.7M). The far lead uppercut into the air did not: turning the hips to the right brings the left shoulder forward, which is the easy way to reach a target 0.19 m further out. Raising the facing term from 0.3 to 1.0 in upL3 changed nothing worth keeping.

The stagger

On upL2-209.7M’s long video I sent a screenshot of the far air uppercut and said: “It lost balance here.” After the punch he hopped on one leg with the feet nearly in a line, and the gate had counted no fall. The gate now counts steps after the punch and the highest foot lift. Measured that way, every lead-uppercut model staggered on the far air rows (upL2-209.7M 1–2 steps, a foot up 7–9 cm), and the student distill4 was the worst (2–3 steps, 9–17 cm). Nothing staggered at the bag, and the rear uppercut never did. With nothing to hit, the reach carried the body forward, and only a quarter of the presses had no target.

I chose the fix: a reward for keeping the feet down once the punch is over, and more of the far uppercut into the air, with the reach left at 0.19 m. That was upL4, and it fixed more than the feet. Every snapshot planted after the far air uppercut (0 steps, foot within about 1 cm), and the far end, flat since upL1, was finally learnt: 54/54 air hits. Bag hits stayed at 0/54, as for every lead uppercut so far, because it reaches the bag at about 4 m/s, under the hit speed.

upL4’s first pod was terminated by the launcher after a single failed pull from the network volume (“Connection reset by peer”) while the trainer was fine. The launcher now logs a failed pull mid-run and lets the next one catch up. The run resumed on a second pod and finished at 170.4 M.

What came next

upL4-157.3M and upR2-209.7M were kept provisionally, pending my look at the videos, and went straight into the student distill7. When I passed distill7 (“Looks good”), the two teachers stopped being provisional.

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