Mutual Alignment Transfer Learning

About

Training robots for operation in the real world is a complex, time consuming and potentially expensive task. Despite significant success of reinforcement learning in games and simulations, research in real robot applications has not been able to match similar progress. While sample complexity can be reduced by training policies in simulation, such policies can perform sub-optimally on the real platform given imperfect calibration of model dynamics. We present an approach -- supplemental to fine tuning on the real robot -- to further benefit from parallel access to a simulator during training and reduce sample requirements on the real robot. The developed approach harnesses auxiliary rewards to guide the exploration for the real world agent based on the proficiency of the agent in simulation and vice versa. In this context, we demonstrate empirically that the reciprocal alignment for both agents provides further benefit as the agent in simulation can adjust to optimize its behaviour for states commonly visited by the real-world agent.

Markus Wulfmeier, Ingmar Posner, Pieter Abbeel · 2017

Related benchmarks

Task	Dataset	Result
Off-dynamics Reinforcement Learning	Reacher broken source environment MuJoCo	Average Reward30	7
Reinforcement Learning	MuJoCo Reacher 1.5 density v1 (test)	Reward-11.1	7
Reinforcement Learning	Walker2d 1.5 gravity MuJoCo	Reward1419	7
Reinforcement Learning	MuJoCo Walker2d 1.5 density v1 (test)	Reward1498	7
Reinforcement Learning	Walker2d 0.5 gravity (test)	Average Return767	7
Reinforcement Learning	MuJoCo Ant 1.5 density v1 (test)	Reward3135	7
Off-dynamics Reinforcement Learning	HalfCheetah 0.5 density v1 (test)	Reward2676	7
Off-dynamics Reinforcement Learning	Reacher 0.5 density v1 (test)	Reward-13.2	7
Reinforcement Learning	Ant 0.5 gravity (test)	Average Return980	7
Reinforcement Learning	Reacher 0.5 gravity (test)	Average Return-13.6	7