Work log · Experiment · Muscles · Punches · RunPod

Work log for Muscles instead of motors

Muscles, stage 3: the punches

By Tony Yum · 28 October 2026

Training cost about $11.18RunPod RTX 5090, RTX 4090 and RTX PRO 6000 pods, $0.89–2.49 an hour · 10 training runs, 16 machines, about 9 machine-hours

The punch env is the one the five punches were learnt in (jab, cross, both hooks, the uppercut, pressed on command at targets in the air and at a heavy bag, with shoves in between). A punch counts as a hit when the fist passes within 5 cm of the target at 3.5 m/s or faster. Every run starts from its body’s own stance; the servo body is the yardstick. Before the first run, muscle mode went into the punch and survive envs, on CPU and GPU, and the GPU copy was checked against the CPU step by step on both bodies (muscle3check, six pods).

The runs

runbody, startwhat changedresult
jab3sservo, its stancefull shoves145 M steps: falls under 3 %, hits 4 % → 9 %, miss 25 → 10 cm
jab3scservoshoves ramped up (a curriculum)141 M steps: hits 5 % → 8 %; the curriculum adds nothing for the servo body
jab3m2a muscles, muscle2S-33.1Mfull shoves52 % falls, no hits; stopped at 31 M
jab3c2a musclesshove curriculumfalls about 20 %, then a collapse; no hits, misses 51–68 cm (the arm stayed in the guard); stopped at 69 M
jab3r2a muscles+ a stretch reflexthe arm comes out (misses 28–31 cm), but the stance under it was trained without the reflex: falls 25–30 %, no hits; stopped at 39 M
muscle2R2a stance from scratch with the reflexsolved at 56.0 M steps, pose 0.93 (kept)
jab3R2a, muscle2Rthe reflex all the wayfalls about 22 %, no hits, fists 0.5–2.9 m/s; stopped at 42 M
muscle2Arelaxed arms: arm muscles rest at 10 %, not 50 %the stance from scratch with the reflexsolved at 47.1 M steps, pose 0.92 (kept)
jab3Arelaxed arms, muscle2Afirst muscle hit at 3.6 M steps (an uppercut, 4.3 m/s), then slow, safe punches to 133 M: falls about 24 %, hits 0.1–0.6 %
jab3Nrelaxed arms, muscle2Aa third of the exploration noisefalls about 95 % from the start; stopped at 22 M

What it showed

  • The muscle punches were too slow to count. jab3R’s fists moved at 0.5–2.9 m/s against the servo body’s 3–11. Both muscles of every arm pair rested at half excitation. With the jab’s muscles forced on and nothing else changed, the fist reached 8.6 m/s with the opposing muscles off and 3.0 m/s with them at half: they brake the arm.
  • Relaxed arms fixed the braking, not the punching. With the arm muscles resting at 10 %, the opposing muscles switch off during a punch and the first muscle hit came within 4 M steps. Then the policy chose slow punches that keep it on its feet, and stayed there. Without exploration noise it fell 0 times in 20; with it, about a quarter of episodes.
  • Stopped here: the next step is a bigger build (a teacher for the arms, or an easier speed bar to begin with), not more steps.

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
jab3N-2026102828 OctoberStopped10.4$0.33
jab3A-2026102828 OctoberNot kept22.1$1.86
muscle2A-2026102828 October · kept muscle2A-47.1MKept–0$0
jab3R-2026102828 OctoberStopped10.5$1.35
muscle2R-2026102828 October · kept muscle2R-56.0MKept10.5$1.32
jab3r-2026102828 OctoberStopped10.6$0.49
jab3sc-2026102828 OctoberNot kept11.5$1.84
jab3c-2026102828 OctoberStopped10.8$1.01
jab3s-2026102828 OctoberNot kept11.6$1.85
jab3m-2026102828 OctoberStopped10.4$0.43
10 training runs108.4$10.49
  • muscles stage 3 shakedown: GPU parity of the survive and punch envs in muscle mode (and the servo body unchanged), a 3-iteration skill-trainer smoke run, 28 October: 6 machines, 0.6 h, $0.69

← All the work logs for “Muscles instead of motors”