Dynamics Learning with Cascaded Variational Inference for Multi-Step Manipulation

Kuan Fang, Yuke Zhu, Animesh Garg, Silvio Savarese, Li Fei-Fei
; Proceedings of the Conference on Robot Learning, PMLR 100:42-52, 2020.

Abstract

The fundamental challenge of planning for multi-step manipulation is to find effective and plausible action sequences that lead to the task goal. We present Cascaded Variational Inference Planner (CAVIN), a model-based method that hierarchically generates plans by sampling from latent spaces. To facilitate planning over long time horizons, our method learns latent representations that decouple the prediction of high-level effects from the generation of low-level motions through cascaded variational inference. This enables us to model dynamics at two different levels of temporal resolutions for hierarchical planning. We evaluate our approach in three multi-step robotic manipulation tasks in cluttered tabletop environments given raw visual observations. Empirical results demonstrate that the proposed method outperforms state-of-the-art model-based approaches by strategically planning for interactions with multiple objects. See more details at pair.stanford.edu/cavin

Cite this Paper


BibTeX
@InProceedings{pmlr-v100-fang20a, title = {Dynamics Learning with Cascaded Variational Inference for Multi-Step Manipulation}, author = {Fang, Kuan and Zhu, Yuke and Garg, Animesh and Savarese, Silvio and Fei-Fei, Li}, pages = {42--52}, year = {2020}, editor = {Leslie Pack Kaelbling and Danica Kragic and Komei Sugiura}, volume = {100}, series = {Proceedings of Machine Learning Research}, address = {}, month = {30 Oct--01 Nov}, publisher = {PMLR}, pdf = {http://proceedings.mlr.press/v100/fang20a/fang20a.pdf}, url = {http://proceedings.mlr.press/v100/fang20a.html}, abstract = {The fundamental challenge of planning for multi-step manipulation is to find effective and plausible action sequences that lead to the task goal. We present Cascaded Variational Inference Planner (CAVIN), a model-based method that hierarchically generates plans by sampling from latent spaces. To facilitate planning over long time horizons, our method learns latent representations that decouple the prediction of high-level effects from the generation of low-level motions through cascaded variational inference. This enables us to model dynamics at two different levels of temporal resolutions for hierarchical planning. We evaluate our approach in three multi-step robotic manipulation tasks in cluttered tabletop environments given raw visual observations. Empirical results demonstrate that the proposed method outperforms state-of-the-art model-based approaches by strategically planning for interactions with multiple objects. See more details at pair.stanford.edu/cavin} }
Endnote
%0 Conference Paper %T Dynamics Learning with Cascaded Variational Inference for Multi-Step Manipulation %A Kuan Fang %A Yuke Zhu %A Animesh Garg %A Silvio Savarese %A Li Fei-Fei %B Proceedings of the Conference on Robot Learning %C Proceedings of Machine Learning Research %D 2020 %E Leslie Pack Kaelbling %E Danica Kragic %E Komei Sugiura %F pmlr-v100-fang20a %I PMLR %J Proceedings of Machine Learning Research %P 42--52 %U http://proceedings.mlr.press %V 100 %W PMLR %X The fundamental challenge of planning for multi-step manipulation is to find effective and plausible action sequences that lead to the task goal. We present Cascaded Variational Inference Planner (CAVIN), a model-based method that hierarchically generates plans by sampling from latent spaces. To facilitate planning over long time horizons, our method learns latent representations that decouple the prediction of high-level effects from the generation of low-level motions through cascaded variational inference. This enables us to model dynamics at two different levels of temporal resolutions for hierarchical planning. We evaluate our approach in three multi-step robotic manipulation tasks in cluttered tabletop environments given raw visual observations. Empirical results demonstrate that the proposed method outperforms state-of-the-art model-based approaches by strategically planning for interactions with multiple objects. See more details at pair.stanford.edu/cavin
APA
Fang, K., Zhu, Y., Garg, A., Savarese, S. & Fei-Fei, L.. (2020). Dynamics Learning with Cascaded Variational Inference for Multi-Step Manipulation. Proceedings of the Conference on Robot Learning, in PMLR 100:42-52

Related Material