One Policy to Control Them All - CrossMinds.ai
One Policy to Control Them All
Jul 14, 20207 views
Deepak Pathak
This video summarizes our ICML 2020 paper on "One Policy to Control Them All: Shared Modular Policies for Agent-Agnostic Control". More details, source code, and paper can be found at https://huangwl18.github.io/modular-rl/

TL;DR: Deep learning success in CV/NLP is driven by the ability to pretrain on diverse data. How can we bring pretraining success to robots? Each robot has different shapes, actions, motors, etc. Furthermore, vanilla RL doesn't generalize beyond the training agent. In this project, we train a single policy to control wide variety of planar robots and then show generalization to unseen robotic agents at test-time without any further training.

Project Abstract:
Reinforcement learning is typically concerned with learning control policies tailored to a particular agent. We investigate whether there exists a single global policy that can generalize to control a wide variety of agent morphologies -- ones in which even dimensionality of state and action spaces changes. We propose to express this global policy as a collection of identical modular neural networks, dubbed as Shared Modular Policies (SMP), that correspond to each of the agent's actuators. Every module is only responsible for controlling its corresponding actuator and receives information from only its local sensors. In addition, messages are passed between modules, propagating information between distant modules. We show that a single modular policy can successfully generate locomotion behaviors for several planar agents with different skeletal structures such as monopod hoppers, quadrupeds, bipeds, and generalize to variants not seen during training -- a process that would normally require training and manual hyperparameter tuning for each morphology. We observe that a wide variety of drastically diverse locomotion styles across morphologies as well as centralized coordination emerges via message passing between decentralized modules purely from the reinforcement learning objective.

Music by Vincent Rubinetti (courtesy @3Blue1Brown)
Download the music on Bandcamp:
https://vincerubinetti.bandcamp.com/album/the-music-of-3blue1brown
Stream the music on Spotify:
https://open.spotify.com/album/1dVyjwS8FBqXhRunaG5W5u
ICML 2020
Recommended