Teacher: the dash

A dash is a quick push-off forward, back or to the side, landing back in the stance. I wanted one in any direction and at any distance, shorter and longer than my own, with the diagonals mixing the real dashes either side. The fast stepping of 22 October was stopped because its scripted references never got the body to push off and travel, so I redid the references from motion capture (Muay Thai 03, take_02): my forward dash at 8.4 s, back at 9.2 s, left at 18.65 s and right at 19.45 s. From these one table was built: 8 directions every 45° by 5 sizes up to 0.8 m, the legs by inverse kinematics. The dash has its own button and its own clock, and a landing counts as ok when it travels within 15 cm of the distance, drifts less than 5 cm sideways and ends in a stance within 10 cm of the right length. The technique is on the dash page.

The runs

Each run was warm-started from the last, all on RunPod in EUR-IS-1, with distill4-131.1M as the teacher for everything but the dash.

run	from	change	outcome	machine-hours, cost
dash1	distill4-131.1M	the dash, in the opponent env with every other skill	189.1 M at its 2 h cap: ok 0.24, about 23 cm short	2.1 h, $2.47
dash1b	dash1-189.1M	every episode a dash episode	ok 0.37–0.39, 12–15 cm short; failed on video: bows forward	2.7 h, $2.08
dash1c (first try)	dash1b-173.7M	a lean term, measured wrongly	stopped at 1.8 M	0.3 h, $0.31
dash1c	dash1b-173.7M	the chest’s lean from vertical held to the reference’s	222.8 M: the bow mostly gone, the dash held	2.5 h, $2.64
dash1d	dash1c-222.8M	the hips’ drop held to the reference’s	108.1 M: mixed; back-right still weak	1.3 h, $1.37
dash1e	dash1d-108.1M	each direction asked at most what dash1d covered	kept 85.2M: 5/8 ok at each direction’s maximum, 81–109 % of the distance	1.2 h, $1.05

The costs include the video pods. dash1b’s first two video pods failed on a bug in the new video script.

Distance first

dash1 trained from 05:22 UTC on 26 October. Half its episodes were the opponent env’s: blocks, punches, walking. It ended landing ok 24 % of the time, still improving slowly. I said “there’s no need to do any blocking training”: the teacher only needs the dash, and the student gets the rest from the other teachers. I also said: “It’s important that it’s stepping far and also fast”. So dash1b made every episode a dash episode and kept the dash’s own timing. That helped (ok 0.39, short about 13 cm). At the end the network volume was full, and the final checkpoint was written as 0 bytes, so 173.7 M, the last good snapshot, is the one that was measured.

Its video showed the short dash working (6/8 ok at 0.4 m) and the long one not. At 0.8 m it still covered only about 0.4–0.6 m, as if it had one dash size.

Form: the bow and the squat

I judged dash1b’s distance “okay for now. But then form is wrong. it shouldn’t bend forward”. The chest was about 45° forward with the head down, where my captured dash keeps it 0–6° forward. dash1c added a lean term, and the hips’ speed rewards are paid only as far as the body stays upright, so a bow buys no speed. The first try measured the lean along the hips’ heading, which barely sees a sideways bend in a side-on stance, so it was stopped after 1.8 M and relaunched with the total lean from vertical. The bow fell from 9° to about 3.4° (the training average) within 5 M steps. On video the worst direction, back-right, went from 49–58° to 18–19°.

On dash1c’s video I said: “Mostly good, back right is the only one that looks strange. It does a squat”. My captured dash drops the hips at most 4–7 cm, and back-right dropped 12 cm more than that. dash1d added the hips’ drop to the same term. The short back-right squat halved (12 to 5.3 cm), but it then covered only 0.17 m, and the long back-right still squatted and drifted 17 cm sideways.

Realistic distances

On dash1d’s video I asked: “let’s keep those distances as expectations so when training we don’t train unrealistic targets”. Nothing had ever covered 0.8 m, and my own captures are 0.45–0.49 m. dash1e caps each direction at what dash1d covered: from 0.32 m (forward-right) to 0.66 m (back). It trained from 00:13 UTC on 27 October for one hour on an RTX 4090:

direction	asked at most	went	sideways	ok
forward	0.47 m	0.43 m (92 %)	1.2 cm	ok
fwd-left	0.54 m	0.44 m (81 %)	1.0 cm	ok
left	0.44 m	0.48 m (109 %)	2.7 cm	ok
back-left	0.40 m	0.38 m (95 %)	6.7 cm	–
back	0.66 m	0.61 m (93 %)	6.2 cm	–
back-right	0.41 m	0.38 m (93 %)	21.0 cm	–
right	0.53 m	0.47 m (89 %)	1.1 cm	ok
fwd-right	0.32 m	0.31 m (98 %)	1.3 cm	ok

In training it landed ok 59–67 % of the time (dash1d about 38 %) and fell short by 3–4.5 cm. The average lean stayed about 3.3°, and it never fell. On video the squat was at most 5.4 cm. The back three directions miss on sideways drift, not distance. I said: “Looks good upload the model to S3”. dash1e-85.2M is the dash teacher. Back-right is the weakest in every run. It is a mix of my back and right dashes, and a captured back-right dash is the likely fix. The dash is not in the 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