# Raw notes "The field of embodied cognition posits that intelligent behaviors can be rapidly learned by agents whose morphologies are well adapted to their environment" 3-5 9,10 - embodied agents Difficulties - morphologies have large search spaces; leading to compromises: - limited morphlogical search spaces 11-17 - hand-designed morphology 17-20 - lifetime learning to evaluate fitness of a morph + control needs lots of computation; leading to compromises: - learn control from something over than raw sensors (11-13, 16, 21-24) - learn <=100 parameters of hand-designed controllers (11-13, 16, 24) - learn to predict the fitness of a morphlogy 15, 21 - mimic Lamarckian evo by giving offspring learned information from parent (12, 15) - limiting the degrees of freedom (13,15) - using cuboids to "simplify the problem" (11-13) 11 - Sims 12 - 2019 - Jelisavcic - Lamarkian Evolution 13 - 2014 - Auerbach, Bongard - Environmental influence on the evolution of morphology complexity 14 - Auerbach - Robogen: Alife 2014 15 - 2019 - Wang, Zhou, Fidler, Ba ICLR ???? (Predict fitness) 16 - Miras, De Carlo, Akhatou, Eiben - Apps of Evo 17 - Liao 2019 18 - Luck, et al. Data-efficient co-adaptation of morphology and behavior with deep RL 21 - Robogrammar 22 - Cheney, MacCurdy, Clune, Lipson, Unshackling Evoliution 23 - Cheney Bongard, SunSpiral, Lipson. Scalable co-optimization of morph and control in embodied machines. Key Claimed Contributions - eval morph by measuring speed to learn a suite of tasks - role of environmental complexity - demonstrate morphological Baldwin effect - evolution rapidly selects morphs that learn faster, enabling behaviors learned late in life of early ancestors to be expressed early in lifetime of descendants. - "mechanistic" foundation that morphs that are physically stable and energy efficient are better suited for learning ## Approach - Nero style learning (?) - "lifetime learning using RL" (34) - tournaments in groups of four, mutated copy added to the pop - PPO - 1152 CPUs over 16 hours ### UNIMAL - parameterization of morphology Morphology is represented by a kinematic tree corresponding to a hierarchy of 3d rigid parts. Parts are connected by motor-actuated hinge joints. Parts are either spheres (the root/head) or capsules (limbs) ### Mutations