Behavioral Repertoires for Soft Tensegrity Robots

Doney, Kyle; Petridou, Aikaterini; Karaul, Jacob; Khan, Ahsan Y.; Liu, Geoffrey; Rieffel, John

doi:10.1109/ssci47803.2020.9308218

Cited by 10 publications

(7 citation statements)

References 35 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…As in the navigation experiments, the middle rows of this figure indicate that QD methods for which no priors are available (generation 0) are not able to solve all of the sampled environments. This effect seems to be much more pronounced for basketball-v2, where a QD method initialised task name QD R [1] QD G [2] FAERY R [3] FAERY G [4] FAERY it [5] c µ [6] λ [7] M 1.0 3.9 1.0 0.0 80 100 100 10 10 1 QD R: percentage of solved environments (normalised in [0, 1]) when using QD optimisation from scratch. 2 QD G: Average number of necessary generations required to solve an environment, when using QD optimisation from scratch.…”

Section: A Few-shot Learning For Single-task Generalisationmentioning

confidence: 99%

“…4 FAERY G: Average number of necessary generations required to solve an environment, when using QD with priors learned by FAERY. 5 FAERY itc: Number of generation at which the prior population has converged in terms of FAERY R and FAERY G. 6,7 µ, λ: respectively population size and number of offsprings. 8,9 Mtrain, Mtest: number of environments used for training and testing.…”

Section: A Few-shot Learning For Single-task Generalisationmentioning

confidence: 99%

“…shelf-place-v2, pick-place-v2 and bin-picking-v2 as the target tasks 7 . The latter three environments had proven to be difficult to solve in a reasonable number of generations using QD optimisation, mainly because of the high reward returned for brittle policies that exploit the particularities and bugs in the physics simulator, making the occurence of proper grasping behaviors rare.…”

Section: B Cross-task Generalisationmentioning

confidence: 99%

“…In part because of their natural ability to deal with those issues, Quality-Diversity [1] (QD) methods have garnered considerable interest in the past few years, and have been applied to many problems, ranging from robotics [2], [3], [4], [5], [6], [7] to video games [8], [9], [10] and open-ended learning [11], [12]. The aim of these methods is to obtain a diverse set of well-performing solutions to a given problem.…”

Section: Introductionmentioning

confidence: 99%

See 3 more Smart Citations

Few-shot Quality-Diversity Optimization

Salehi,

Coninx,

Doncieux

2021

Preprint

View full text Add to dashboard Cite

In the past few years, a considerable amount of research has been dedicated to the exploitation of previous learning experiences and the design of Few-shot and Meta Learning approaches, in problem domains ranging from Computer Vision to Reinforcement Learning based control. A notable exception, where to the best of our knowledge, little to no effort has been made in this direction is Quality-Diversity (QD) optimisation. QD methods have been shown to be effective tools in dealing with deceptive minima and sparse rewards in Reinforcement Learning. However, they remain costly due to their reliance on inherently sample inefficient evolutionary processes. We show that, given examples from a task distribution, information about the paths taken by optimisation in parameter space can be leveraged to build a prior population, which when used to initialise QD methods in unseen environments, allows for few-shot adaptation. Our proposed method does not require backpropagation. It is simple to implement and scale, and furthermore, it is agnostic to the underlying models that are being trained. Experiments carried in both sparse and dense reward settings using robotic manipulation and navigation benchmarks show that it considerably reduces the number of generations that are required for QD optimisation in these environments.

show abstract

Section: A Few-shot Learning For Single-task Generalisationmentioning

confidence: 99%

Section: A Few-shot Learning For Single-task Generalisationmentioning

confidence: 99%

Section: B Cross-task Generalisationmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

Few-shot Quality-Diversity Optimization

Salehi,

Coninx,

Doncieux

2021

Preprint

View full text Add to dashboard Cite

show abstract

“…Many studies have pursued dynamical control methods for tensegrity robots. For example, the use of vibrational motors in controlling spherical tensegrities was effective in achieving high locomotion speeds with respect to body size and finding various translational and rotational movements [15], [16]. Bliss et al used central pattern generators (CPGs) to generate motion for a tensegrity-based swimming robot that can exploit resonance entrainment [17].…”

mentioning

confidence: 99%

Behavioral Diversity Generated From Body–Environment Interactions in a Simulated Tensegrity Robot

Terajima

Inoue²,

Yonekura³

et al. 2022

IEEE Robot. Autom. Lett.

View full text Add to dashboard Cite

Tensegrity structures, which are made of struts and tendons, are attracting attention as a platform for adaptive and resilient robots with connections to biological systems. However, they are difficult to control because of their elasticity, deformability, and tight coupling between their elements. Studies on morphological computation and physical reservoir computing suggest, however, that the temporal patterns generated by body dynamics can be exploited to perform computations. In this work, we analyze the diverse collection of behaviors generated by driving tensegrity robots with simple periodic motor commands. We find that characteristic locomotion gaits, such as sliding and rolling, appear in specific regions of the system parameter space. Furthermore, the analysis shows that both normal and anomalous deterministic diffusion emerge because of interactions of the body with the environment. The highly nonlinear relationship between the parameters and robot behavior highlights the difficulty of controlling tensegrities. However, we demonstrate that our results of these nontrivial relationships can in fact be directly exploited to achieve adaptive behavioral switching. These results point to potential uses of tensegrity dynamics as computational resources.

show abstract