Fig. 4: Ablation experiments demonstrate the general applicability and minimal impact of curriculum schedules.
From: Bioinspired morphology and task curricula for learning locomotion in bipedal muscle-actuated systems
We compare the ontogenetic morphological and task curriculum (OntoCurr), the uniform morphological and task curriculum (UniCurr), only the task curriculum (AdultCurr), and random morphology schedules with uniform (UniRand) and ontogenetic (OntoRand) scaling. We test five curriculum schedules, differentiated by curriculum and morphology switch points (dotted and dashed lines, respectively), and present the corresponding means and standard deviations of rewards. a Schedule 2-4-8 includes 2 × 106 curricula increments and is complemented by a non-curriculum baseline employing state-of-the-art DEP-RL19. To maintain clarity, the AdultCurr condition, identical to the schedule in the main results, is not presented. b Schedule 3-6-12 assigns 3 × 106 step increments. c Schedule 4-8-16 gives 4 × 106 step increments. d Schedule 2-4-8-12 builds on (a) allocating steps 8 × 106 to 12 × 106 to the corresponding 12 year-old morphology before switching to the adult morphology and is thus only presented for double curricula strategies. e The fifth schedule abandons both morphology and task curricula, opting for random switches in both morphology and velocity.
