Teachers: the elbows

The first elbows, both from my motion capture (Muay Thai 03, take_03). For the right elbow I picked the one at “6 s”: the rear foot steps in about 35 cm, the elbow sweeps right to left across the face at head height as the torso turns about 110°, then the rear foot slides back into a high guard. I asked for it to land on the chin and to be trained at the points around it, and left those to the AI: “you decide on all the points around the aim”. The left elbow is the one at 13 s, a stepping elbow: the lead foot steps in about 0.4 m, the rear foot skips after it, the elbow sweeps left to right, and he steps back to where he started. I asked for separate teachers for the two, trained side by side. The technique is on the elbows page.

Before training: fixing the capture

The take needed work first. Where my feet went off the bottom of one phone’s picture they were re-placed from the other phone. Without that the rear leg locked straight and the lead foot didn’t step. Then, from my notes, the right elbow was lifted to the chin (1.49 m; it peaked at 1.42 in the capture), the left glove was put against the left ear as the elbow lands, and the elbow was tucked in afterwards. The aims are the chin ±7.5 cm up, down, left and right, and ±6 cm nearer or farther. The arm and the spine are solved onto each aim by inverse kinematics. Inside that range the right elbow’s reference lands within 0.6 cm of its target. The left one is harder. I drop about 9 cm into the step and my elbow lands at 1.38 m, 11 cm under the chin, so the spine takes more of the reach: it lands within 2.6 cm inside the presses. A one-frame shoulder flip in the retarget at 14.03 s was smoothed over 0.3 s.

The runs

run	from	what	outcome	machine-hours, cost
elbowR1	distill4-131.1M	the right elbow to the chin	training done: elbowR1-249.1M, taken into distill8 before its own gate	4.5 h, $3.46
elbowL1	distill4-131.1M	the left, stepping elbow	training done: elbowL1-256.7M, taken into distill8 before its own gate	5.0 h, $3.38

Both were launched at 14:17–14:18 UTC on 26 October, each on a rented RTX 5090 on vast.ai with a 2.5-hour training cap. They were warm-started from distill4-131.1M, with distill4 as the teacher for everything but the elbow. distill7 had not passed its gate yet. Getting the machines took a while. The right elbow’s first box was still loading after 10 minutes and was dropped. The left elbow went through four boxes (still loading, or no ssh) before the fifth worked, about 40 minutes after the first try. Those boxes are in the costs.

What is known so far

The run files have no outcome written yet. What there is comes from distill8’s log. At the end of training the right elbow hit 25 % of its presses, missing by 7.1 cm on average. The skills check on the boxes after training showed a side effect: the elbow teachers passed the opponent only 6 and 15 times out of 160, with 145–149 falls. They were trained with no opponent episodes. That matters little for a teacher, which only acts while its own technique runs. In distill8 the opponent score held at 112/160.

On 27 October I asked to start distilling both elbows into the student without waiting for their gate. This log is still going: each teacher’s own gate (the chin and the eight aims around it, in the air and at the bag, the turn back after the elbow, the feet at the end) started alongside distill8’s, and the videos come after it.

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