An Invitation to Imitation

By Andrew Bagnell, March. 2015

91 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

A Survey of Deep Network Solutions for Learning Control in Robotics: From Reinforcement to Imitation (2016, neutral): “Imitation learning, as an alternative to learning control policies, guides the policy search, not by hand-designed reward signals, but by providing the learning agent with experts’ demonstrations (Bagnell, 2015).”
Learning Self-Correctable Policies and Value Functions from Demonstrations with Negative Sampling (2019, background): “Behavioral Cloning [Bain and Sommut, 1999]. Behavioral Cloning (BC) learns a mapping from a state to an action on demonstration data using supervised learning. We use MSE loss for predicting the actions. Nair et al.’18 [Nair et al., 2018]. The previous state-of-the-art algorithm from Nair et al. [20”
Gated-Attention Architectures for Task-Oriented Language Grounding (2017, methodology): “We consider two different learning approaches for language grounding using target-driven visual navigation: (1) Imitation Learning (Bagnell, 2015): where the agent has access to an oracle which speciﬁes the optimal action given any state in the environment; (2) Reinforcement Learning (Sutton, 1984):”
An Algorithmic Perspective on Imitation Learning (2018, methodology): “…properties including constraints (for example, robot joint limits), dynamic smoothness and stability, or leading to a coherent, multi-step plan [Bagnell, 2015]; 2) the interaction between the learner’s decisions and its own input distribution (an on-policy versus off-policy distinction) , and…”
AirSim: High-Fidelity Visual and Physical Simulation for Autonomous Vehicles (2017, background): “Recently, paradigms such as reinforcement learning [12], learning-by-demonstration [2] and transfer learning [25] are proving a natural means to train various robotics systems.”
Imitation Learning as f-Divergence Minimization (2019, neutral): “Imitation learning (IL) has a long-standing history in robotics as a tool to program desired skills and behavior in autonomous machines [7–10].”
Model-Free Imitation Learning with Policy Optimization (2016, background): “…the assumption of expert optimality acts as a prior on the space of policies, IRL algorithms can allow the learner to generalize expert behavior to unseen states much more effectively than if the learner had tried to produce a policy or value function instead (Ng & Russell, 2000; Bagnell, 2015).”
Policies for Active Learning from Demonstration (2017, methodology): “Due to these differences, applying standard supervised approaches to LfD meets with limited success (Bagnell 2015), and therefore there is a need for specialized approaches that account for these differences.”
A Bayesian Approach to Generative Adversarial Imitation Learning (2018, background): “However, BC is susceptible to compounding error, which refers to small prediction error accumulated over time to a catastrophic level [Bagnell, 2015].”
Deep Local Trajectory Replanning and Control for Robot Navigation (2019, background): “Due to limited space, we encourage readers interested in more details of these different approaches to refer to [18, 20, 21, 23, 27].”
Mo' States Mo' Problems: Emergency Stop Mechanisms from Observation (2019, background): “We specifically study reset mechanisms which are more natural for human operators to provide – in the form of large red buttons, for example – and thus perhaps less noisy than action or value feedback [5].”
Multi-Agent Generative Adversarial Imitation Learning (2018, background): “There is a vast literature on single-agent imitation learning [33].”
Imitating driver behavior with generative adversarial networks (2017, background): “Poorer predictions can cause a feedback cycle known as cascading errors [20].”
Recent Advances in Robot Learning from Demonstration (2020, background): “However, since demonstrations seldom cover all parts of the problem space, the robot is likely to discover scenarios where the input distributions are different from those of the demonstrations (206).”