Evaluation: the rear knee's lean, step by step

No training here: these boxes measured and rendered models that already existed, for Correcting the knee. They ran on EC2 because my development server is a small burstable machine: a local try used 10 seconds of CPU in 9 minutes before I stopped it.

• kneetrace ran one rear knee in the air and one at the heavy bag, six seeds each, deterministic, for four models: the teachers kneeR3-183.5M and kneeR4-183.5M and the students distill4-131.1M and distill7-245.8M. At every control step (50 a second) it logged the pelvis’s and the chest’s lean, and the corrected reference’s. Its lean-past-the-reference numbers match the gate’s exactly (kneeR3 12.4° / 9.7°, kneeR4 1.3° / 0.9°, distill7 1.4° / 1.2°), which checks the trace against the gate. The new number is distill4’s: 11.4° / 10.7°.
• kneestills rendered the same knee at the bag, seed 0, from behind and from his right side, every 60 ms. The bag didn’t show: its geoms are the target’s, which the script made transparent.
• kneestills2 rendered four moments again with the bag shown: the impact, then 0.3, 0.6 and 0.9 s later.
• kneestills3 rendered the student before and after the fix (distill4-131.1M, distill7-245.8M), six seeds each, from cameras that turn with his hips: from straight behind, and from behind each shoulder, where the line of his back shows side-on. The earlier cameras were fixed in the room, so once his hips had turned 70° neither showed his back. It also logged where each foot was relative to his centre of mass, and how high each foot was. Before the fix his rear foot comes back up 5–11 cm after it lands; after the fix, 4–5 cm. In seed 1 both of his feet are up at once for 0.44 s. His centre of mass sits at the same place between his feet in both models.

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 for “Correcting the knee”