Work log · Teacher · Evasion · Failed models
Work log for Shadow clones: four Muay Thai techniques learnt in parallel
Teachers: slips, rolls and the lean back
Training cost about $12.979 rented RTX 5090s on vast.ai, four at a time, 25 October 2026 · 9 training runs, 11 machines, about 20 machine-hours
Four techniques from one take of my motion capture (Muay Thai 02, take_01): the slip uppercut (slip left, lead uppercut), the lean back, and rolling under to the right and to the left. Each has its own input in opponent_evade, and each teacher is warm from the kept student distill4-131.1M, which holds every other skill. An evasion has no target. Its “hit” is the head crossing within 5 cm of the reference’s head at the apex, at any speed.
Round one (evade1): four boxes at once
| teacher | gate | verdict |
|---|---|---|
| rollR1-238.5M | 8/8, 2.5 cm, no falls | failed on video: drops rather than rolls; runs my clip at 1.24× on average |
| rollL1-226.3M | 8/8, 6.3 cm | failed: as rollR1, and its head sweeps the wrong way across |
| slipup1-204.5M | 8/8, ball 8/8 | kept for now: “the uppercut still looks good, so keep that one”, “but actually even that one looks like a big uppercut than a slip” (1.42× on average, the uppercut ~0.5 s early) |
| leanback1-231.8M | 0/8, 38.9 cm | failed: never leans; the head drops 7 cm at most |
Cause: the punch clock. The policy may play a reference up to twice as fast and is paid for reaching the target sooner. That was made for punches, and only the jab is held to its natural pace. The evasion teachers squeezed my moves, and the gate doesn’t check timing.
Round two (evade2): a wider head path
I asked for the rolls to “roll to the side like a U shape”. The reference was widened (the head 25 cm out), and the head was tracked along its path with a reward.
| teacher | gate | verdict |
|---|---|---|
| slipup2-196.6M | 8/8, 1.9 cm | failed: “looks more like a football headbutt, it just stays there”; the trunk bends sideways 52° |
| rollL2b-183.5M | 8/8, 7.4 cm, the U the right way | failed: trunk side bend 51°, head tilt 59° |
| rollR2-229.9M | 4/8, 10.1 cm | failed: “going sideways but not by bending the torso”; skips the lead-in, ~0.3 s early |
| leanback2-234.4M | 0/8, 19.3 cm | leans now but never reaches the apex |
rollL2’s first box ran at 11,000 steps a second against 30,000–34,000 on the others. It was stopped after 0.56 h ($0.38), the machine was added to a skip list, and the run restarted as rollL2b.
Cause: the edited reference. The pelvis couldn’t move the head 25 cm, so the trunk did it with a side bend at its 50° limit, and the teachers learnt the bend. A rebuilt version, with the pelvis moved and the trunk straight, looked physically wrong to me and was never trained. My decision: “Keep to my original reference, like the one from motion capture.”
Where it stands
No roll or lean back is kept. slipup1 stays the slip-uppercut teacher until a better one exists. Its successor, slipup3, goes back to the original capture with the target at the chin (“the target should be where the chin is, not higher than that”); it is not in the ledger yet. Next for the rest: the punch clock held to the capture’s pace for evasions, and a gate that measures timing and shape, not only the crossing.
| machines | machine-hours | cost | |
|---|---|---|---|
| evade1 (4 teachers) | 4 | 8.67 | $5.56 |
| evade2 (4 teachers, plus rollL2’s slow box) | 5 | 9.06 | $5.90 |
| total | 9 | 17.73 | $11.46 |
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).
| Run | Outcome | Machines | Hours | Cost |
|---|---|---|---|---|
leanback1-2026102524 October · lean back | Failed | 1 | 2.1 | $1.14 |
leanback2-2026102525 October · lean back | Failed | 1 | 2.1 | $1.28 |
rollL1-2026102524 October · roll left | Failed | 1 | 2.2 | $1.56 |
rollR1-2026102524 October · roll right | Failed | 1 | 2.2 | $1.38 |
rollR2-2026102525 October · roll right | Failed | 1 | 2.1 | $1.41 |
rollL2b-2026102525 October · roll left | Failed | 2 | 2.7 | $1.81 |
slipup1-2026102524 October · slip and uppercut · kept slipup1-204.5M | Kept | 1 | 2.2 | $1.48 |
slipup2-2026102525 October · slip and uppercut | Failed | 1 | 2.1 | $1.41 |
slipup3-2026102525 October · slip and uppercut | Not kept | 2 | 2.3 | $1.51 |
| 9 training runs | 11 | 20 | $12.97 |
← All the work logs for “Shadow clones: four Muay Thai techniques learnt in parallel”