Techniques · Movement

The dash

A quick push-off in any direction: in, out or to the side, as far as I'd really go in one step.

Its own training$9.916 runs, 10.1 h
Shared training it was part of$00 runs with other techniques

Where it stands

TechniqueTeacher keptIn the studentOwn runs
dashdash1e-85.2M27 October 2026not yet6

The dash had a false start: 26 short runs on 22 October tried fast steps from a scripted clip of my feet, and the long step never learnt to push off the back leg (work log). I gave up and had the references redone from better motion capture. That became Muay Thai 03, take_02: dashes forward (8.4 s), back (9.2 s), left (18.65 s) and right (19.45 s). From those four, one table covers eight directions at any distance up to 0.8 m, each diagonal mixing the two captures either side of it. The dash has its own button.

dash1 (26 October), cloned from distill4 with every other skill in the mix, landed a dash right 24 % of the time, about 23 cm short (work log). dash1b trained on dash episodes only: right 39 % of the time, 12–15 cm short. Its video showed two things. The short dash worked (6 of 8 directions at 0.4 m), but at 0.8 m it covered the same 0.4–0.6 m, as if it had one dash size. And it bowed forward at the waist, where my capture keeps the chest within 6° of upright. Bowing moved the hips sooner, and the hips’ speed paid far more than the trunk (reward hacking).

dash1c paid for an upright chest, and paid for the hips’ speed only when upright. The bow mostly went (back-right from 49–58° to 18–19°). I then saw back-right squat, and dash1d added a term for the hips’ drop. The short back-right squat halved, but it then covered only 0.17 m. Nothing had ever covered 0.8 m, and my own captured dashes are 0.45–0.49 m, so dash1e asks each direction for at most what dash1d covered (0.32–0.66 m). At each direction’s maximum it is right in 5 of 8 directions and covers 81–109 % of the distance. I kept dash1e-85.2M on 27 October.

The dash isn’t in the student yet. The three backward directions drift sideways (back-right 21 cm at its maximum), and back-right, which is a mix of my back and right dashes, is the weakest in every run. A captured back-right dash is the likely fix.

What I think, as a coach

When the fast stepping failed: “none of this is working. Give up … I’m going to redo everything from scratch using a better motion capture.”

For dash1b: “there’s no need to do any blocking training”, and “It’s important that it’s stepping far and also fast.”

On dash1b’s video, the distance was “okay for now. But then form is wrong. it shouldn’t bend forward”. On dash1c’s: “Mostly good, back right is the only one that looks strange. It does a squat”. On dash1d’s: “let’s keep those distances as expectations so when training we don’t train unrealistic targets”.

Every attempt

Every run that trained it on its own, in order.

RunOutcomeMachinesHoursCost
dash1-2026102626 October · dash · kept dash1-189.1M · Teacher: the dashKept12.1$2.47
dash1b-2026102626 October · dash · Teacher: the dashFailed42.7$2.08
dash1c-2026102626 October · dash · Teacher: the dashStopped10.3$0.31
dash1c-2026102726 October · dash · kept dash1c-222.8M · Teacher: the dashKept32.5$2.64
dash1d-2026102726 October · dash · kept dash1d-108.1M · Teacher: the dashKept21.3$1.37
dash1e-2026102727 October · dash · kept dash1e-85.2M · Teacher: the dashKept21.2$1.05
6 training runs1310.1$9.91

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.

Model lineage: 7 runs, each an arrow from the run it started from, dashed from its teachers.Walking and steppingStudents24 Oct26 Oct27 Octdistill4dash1dash1bdash1cdash1cdash1ddash1e
keptstudent (distilled)not kept, stopped or failedstarted fromlearnt from (teacher)Tap a run for its details. Scroll sideways for time.

The work logs