Teacher: the rear knee, landing upright

The rear knee from the first knee teachers, kneeR3-183.5M, landed well, but I noticed: “after the knee and the feet land back on the ground the fighter arches its back”. It was following my motion capture too faithfully. After the impact I lean back a little (pelvis 5–9° back, chest 10–16°, against the stance’s 0 and +1.5°), and the model tipped its pelvis up to about 8° further back still and bent its lower back about 20° to keep the chest on the reference: hips forward, chest back. The technique is on the knees page.

What changed

• The reference. Over 7.05–7.35 s of the capture, just after the hip thrust, the pelvis’s lean and the spine’s bend are brought to the stance’s and held there to the end. The reference is identical before 7.05 s, and the impact (6.933 s) is unchanged. I approved the preview before training. The switch knee and the body straight were not touched.
• A new reward term. After the hit, a term (weight 1.5, a 5° Gaussian) that pays for the chest and pelvis not leaning further back than the reference; leaning forward of it is free.
• The gate adds a row: the most the body leans back past the reference after the hit.

Against the new reference, kneeR3-183.5M leaned back 12.4° in the air and 9.7° at the bag.

The run

run	from	change	outcome	machine-hours, cost
kneeR4	kneeR3-183.5M	the upright reference and the lean-back term	kept 183.5M: lean back 1.3° in the air, 0.9° at the bag; bag 6/6, no falls	2.1 h, $1.29

It ran on one rented RTX 5090 on vast.ai (Pennsylvania, $0.601/h) from 00:03 UTC on 26 October, with kneeR3-183.5M as the warm start and guard14-45.9M as the teacher, as for kneeR3. It reached its 2-hour cap at 207.9 M steps, at about 30,000 steps a second. The lean dropped almost at once: by 13.1 M the bag row was at 1.8° (kneeR3 9.2°).

The gate

Six seeds per snapshot, the lean measured from the start of the upright window:

snapshot	lean back, air / bag	note
kneeR3-183.5M	12.4° / 9.7°
52 M	5.1° / 1.0°	air hits 5/6
105 M	1.2° / 1.0°	1/6 falls under shoves
157 M	1.4° / 1.4°	the feet end turned in the air
183.5 M	1.3° / 0.9°	feet back in the stance, 0 falls
206 M	3.1° / 0.6°	up to 10.8° in the air

183.5 M was the cleanest: bag 6/6, the feet back in the stance at the end (lead foot 0.40–0.41 m ahead, rear foot where it started), no falls. In the air it hits 0/6, as kneeR3 did. The knee arrives 3.6 cm from the target at 2.5 m/s, under the speed the gate counts as a hit. The clip went to me, and kneeR4-183.5M replaced kneeR3 as the rear-knee teacher. Next it went into the student, in distill7.

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