Machine learning of motor skills for robotics

By Stefan Schaal, Jan Peters. 2007

40 citations recorded by the source corpus.

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

Citations in the corpus

Learning Complex Dexterous Manipulation with Deep Reinforcement Learning and Demonstrations (2017, methodology): “A number of pre-conditioned policy gradient methods have been developed in literature [19], [4], [35], [34], [43], [40], [44] and in principle any of them could be used.”
Learning Policy Improvements with Path Integrals (2010, background): “For this kind of direct policy learning, a general cost function J = ∫ τ p(τ )R(τ )dτ is usually assumed (Peters, 2007) and optimized over state xt and action at trajectories τ = (xt0 ,ato , .”
On a Connection between Importance Sampling and the Likelihood Ratio Policy Gradient (2010, background): “Natural actor critic (NAC) approaches have enjoyed substantial success on real-life robotics tasks [1, 23].”
Learning to grasp under uncertainty (2011, result): “This generic loop is similar to other reinforcement learning algorithms [ 14 ].”
Stochastic search using the natural gradient (2009, methodology): “However, baselines as found in (Peters, 2007) are suboptimal w.r.t. the variance of d., since this FIM may not be invertible.”
A Generalized Path Integral Control Approach to Reinforcement Learning (2010, methodology): “Alternatively, direct policy learning from trajectory roll-outs has recently made significant progress (Peters, 2007), but can still become numerically brittle and full of open ∗.”
Natural Evolution Strategies (2008, background): “The solution to the constrained optimization problem in Equation (5) can be found using a Lagrangian multiplier (Peters, 2007), yielding the necessary condition Fδθ = β∇θJ, for some constant β > 0.”
Reinforcement learning of motor skills in high dimensions: A path integral approach (2010, background): “Alternatively, direct policy learning from trajectory roll-outs has recently made significant progress [1], but can still become numerically brittle and full of open tuning parameters in complex learning problems.”
Reinforcement learning for non-prehensile manipulation: Transfer from simulation to physical system (2018, unknown): “( k) ; (3) where F k is the Fisher Information Metric at the current iterate k. We apply a normalized gradient ascent procedure, which has been shown to further stabilize the training process [38], [39], [4]. This results in the following update rule: k+1 = k+ s gTF 1 k g F 1 k g: (4) The ver”