Work log · Footwork · Failed models

Work log for Shadow clones: four Muay Thai techniques learnt in parallel

Fast stepping and the step-jab (abandoned)

By Tony Yum · 22 October 2026

Training cost about $9.4935 rented RTX 5090s on vast.ai (two of them costed from their run logs; three more never started and cost nothing), 22 October 2026 · 25 training runs, 35 machines, about 12.6 machine-hours

The first job for the whole environment stack rewritten for the GPU: fast steps in and out of range, from a 1.5 s clip of my feet (one step in, straight back out), and a jab that steps in to reach a far target. The parent and teacher was opp15-14.5M, which knew every earlier skill. Its step skills before training were step-jabs 3/40 and fast steps 9/32.

The runs

Each run was short (about 0.2–0.6 h) and warm-started from the last, with one change:

runswhat changed
step0shakedown: the GPU stack at ~22,000 steps a second (three boxes, ~$0.2)
step1–2fast steps and the step-jab; a wider placement term (27.2 M: step-jabs 5/40, fast steps 11/32)
step3–5a slower step-jab clock, distance curricula, the plant fixed, the fast step on its own button
step6a/b, 7a/bthe teacher’s weight A/B; a no-teacher diagnostic (does anything learn when nothing holds the rest?)
step8–10the reach curriculum, the step-jab paid for reaching, capped at 0.30 m
step11–14sliding not bouncing, long steps only, started partway (reference state initialisation), the drive paid as it happens
step15–18the step in reworked (weight back first), slower long steps, the teacher off through the landing
fs1a fast-step specialist, no teacher, every episode a range episode
step19–20bthe long step pushed off the back leg; speed and distance over style, then as a curriculum

Why it stopped

The long step in never learnt to push off the back leg and travel. The step out only covered its distance at the slowed speed. After step20b I stopped it: “none of this is working. Give up … I’m going to redo everything from scratch using a better motion capture.” The step environment and its tools were removed from the code. The run logs and checkpoints are kept, and stepping will come back with new references.

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), plus the tests and renders this log covers.

RunOutcomeMachinesHoursCost
step0-2026102221 OctoberTest30.7$0.48
step1-2026102221 October · jab · kept step1-27.2MKept31.4$1.01
step2-2026102221 October · jabStopped10.5$0.29
step3-2026102221 October · jab · kept step3-15.7MKept10.4$0.23
step4-2026102221 October · jab · kept step4-34.3MKept10.7$0.56
step5-2026102221 October · jab · kept step5-19.4MKept10.5$0.42
step6a-2026102221 October · jabStopped10.4$0.25
step6b-2026102221 October · jab · kept step6b-12.0MKept10.3$0.27
step7a-2026102222 October · jab · kept step7a-18.4MKept10.5$0.31
step7b-2026102222 October · jabStopped20.4$0.33
step8-2026102222 October · jab · kept step8-21.5MKept10.5$0.38
step9-2026102222 October · jab · kept step9-22.4MKept41.2$0.93
step10-2026102222 October · jab · kept step10-28.5MKept10.6$0.44
step11-2026102222 October · kept step11-12.4MKept10.3$0.24
step12-2026102222 OctoberStopped10.2$0.15
step12b-2026102222 October · kept step12b-10.3MKept10.3$0.25
step13-2026102222 October · kept step13-16.2MKept10.3$0.24
step14-2026102222 October · kept step14-16.1MKept10.4$0.32
step15-2026102222 October · kept step15-17.2MKept10.4$0.38
step16-2026102222 October · kept step16-8.0MKept10.2$0.14
step17-2026102222 October · kept step17-9.5MKept10.2$0.13
step18-2026102222 October · kept step18-22MKept10.5$0.42
step19-2026102222 October · kept step19-4.8MKept10.4$0.32
step20-2026102222 OctoberStopped10.4$0.29
step20b-2026102222 OctoberStopped10.7$0.50
fs1-2026102222 OctoberStopped10.2$0.17
25 training runs3412.5$9.42
  • the CUDA-graph crash test, 22 October (gpu-rewrite): 1 machines, 0.1 h, $0.07

← All the work logs for “Shadow clones: four Muay Thai techniques learnt in parallel”