Lab notebook
The blog tells the story; these are the notes underneath it. Each work log covers one stretch of training, run by run: why I started it, what changed, what the videos showed and what it cost. Some went into a post, some haven't yet, and some were dead ends. Everything I've paid for is in one of them.
28 work logs · 260 training runs · about 457 machine-hours · $401 in all
Foundations
- Muscles, stage 2: real leg muscles
MyoSuite's leg muscles on their real paths, mapped onto the fighter: 26 of them, then all 80. A torque-neutral zero action, a cost for resting on joint limits and smaller PPO updates got both to the stance.
- Muscles, stage 1: joint muscles instead of motors
An experiment I asked for: drive the fighter with 86 joint muscles instead of servos, and measure how much longer the stance takes to learn, if it can be learnt at all.
- The higher guard: retraining every skill onto a new stance
Eighteen runs that moved every skill onto a higher guard, fixed the lead hook and the jab, and produced guard18-124.5M, the base every clone starts from.
- Fast stepping and the step-jab (abandoned)Abandoned
Twenty-six short GPU runs on fast steps in and out of range and a jab that steps in. The long step never pushed off the back leg, and the work was stopped to redo the motion capture.
- Run by run: walking in the stance
Every run behind the post, with its outcome, machines and cost.
- Run by run: balancing in the stance
Every run behind the post, with its outcome, machines and cost.
Techniques
- Evaluation: the rear knee's lean, step by step
Four short EC2 boxes that traced the trunk's lean through the rear knee, before and after the upright fix, and rendered the stills for the post.
- Muscles, stage 3: the punchesAbandoned
The five punches on the muscle body, against the servo body on the same settings. The muscle body never punched fast enough to score; why, a stretch reflex and relaxed arms, and where it stopped.
- Teachers: the kicks
Low, mid and high roundhouses on both sides, from my motion capture: two rounds of teachers, all six kept.
- Teacher: the dash
Six runs to a dash teacher: a quick push-off in any direction, upright, without squatting, over the distances I would really cover.
- Teachers: the elbowsStill going
Two elbow teachers from my motion capture: the right elbow to the chin and the stepping left elbow.
- Teachers: the uppercuts, second round
Four more uppercut teachers: the rear one kept facing square, the lead one still turned or staggered until its feet were held down.
- Teacher: the rear knee, landing upright
One more rear-knee teacher, so the knee lands upright instead of arched back.
- Teachers: the uppercuts over distance
Two uppercut teachers trained in parallel on RunPod. The rear one was kept; the lead one turned 90° after the punch, which traced back to how training rewarded facing.
- Teachers: slips, rolls and the lean back
Nine evasion teachers in two rounds of four parallel boxes. They passed the gate and failed on video: rushed by the punch clock, then bent by an edited reference.
- Teacher: the pull counter
One run on RunPod after four vast.ai boxes never loaded: a lean back and a rear straight to the head, 8/8 in the air and at the bag.
- Teacher: the rear straight to the body
Four runs to a body-straight teacher: a no-low-blow line, the bag at the impact point, then a retrain on the base's slower elbows.
- Teachers: the rear knee and the switch knee
Two knee teachers trained side by side, three runs each: the knees landed first time, the feet and the steps home took two more.
- Run by run: blocking
Every run behind the post, with its outcome, machines and cost.
- Run by run: faster punches
Every run behind the post, with its outcome, machines and cost.
- Run by run: five punches
Every run behind the post, with its outcome, machines and cost.
- Run by run: the jab
Every run behind the post, with its outcome, machines and cost.
Students
- Students: the network-size experiment
The student grown from 2 × 512 to 1024 × 2, 512 × 3 and 1024 × 3 without changing what it does, and trained beside a 512 × 2 control for 200 M steps each: the distances to the teachers, the gate, and the four boxes.
- Distillation: distill7 to distill9
Three more students: the uppercuts over distance and the upright knee in distill7, the elbows in distill8, then the high kicks, the pull counter and the dash in distill9.
- Distillation: six runs to the student distill4
Distilling the body straight, both knees and the pull counter into the base: what each run changed, how the gate picked the student, and what it cost to evaluate.
Infrastructure
- Machines that trained nothing: renders and tests
The rented machines that were not training a model: Blender renders for the shadow-clones video and tests of the RunPod launcher.
- Run by run: two fighters, GPU against CPU
Every run behind the post, with its outcome, machines and cost.
- Run by run: GPU against CPU
Every run behind the post, with its outcome, machines and cost.
Not sorted yet
Runs and machines that no work log covers yet ($63.45): newer than the last update of this notebook. They are counted in the totals and on the Stats page all the same.