Generative Adversarial Imitation Learning (GAIL) provides a promising approach to training a generative policy to imitate a demonstrator. It uses on-policy Reinforcement Learning (RL) to optimize a reward signal derived from an adversarial discriminator. However, optimizing GAIL is difficult in practise, with the training loss oscillating during training, slowing convergence. This optimization instability can prevent GAIL from finding a good policy, harming its final performance. In this paper, we study GAIL's optimization from a control-theoretic perspective. We show that GAIL cannot converge to the desired equilibrium. In response, we analyze the training dynamics of GAIL in function space and design a novel controller that not only pushes GAIL to the desired equilibrium but also achieves asymptotic stability in a simplified "one-step" setting. Going from theory to practice, we propose Controlled-GAIL (C-GAIL), which adds a differentiable regularization term on the GAIL objective to stabilize training. Empirically, the C-GAIL regularizer improves the training of various existing GAIL methods, including the popular GAIL-DAC, by speeding up the convergence, reducing the range of oscillation, and matching the expert distribution more closely.

Imitation learning (IL) aims to learn a policy from expert demonstrations that minimizes the discrepancy between the learner and expert behaviors. Various imitation learning algorithms have been proposed with different pre-determined divergences to quantify the discrepancy. This naturally gives rise to the following question: Given a set of expert demonstrations, which divergence can recover the expert policy more accurately with higher data efficiency? In this work, we propose f-GAIL - a new generative adversarial imitation learning model - that automatically learns a discrepancy measure from the f-divergence family as well as a policy capable of producing expert-like behaviors. Compared with IL baselines with various predefined divergence measures, f-GAIL learns better policies with higher data efficiency in six physics-based control tasks.

Consider learning a policy from example expert behavior, without interaction with the expert or access to a reinforcement signal. One approach is to recover the expert's cost function with inverse reinforcement learning, then extract a policy from that cost function with reinforcement learning. This approach is indirect and can be slow. We propose a new general framework for directly extracting a policy from data as if it were obtained by reinforcement learning following inverse reinforcement learning. We show that a certain instantiation of our framework draws an analogy between imitation learning and generative adversarial networks, from which we derive a model-free imitation learning algorithm that obtains significant performance gains over existing model-free methods in imitating complex behaviors in large, high-dimensional environments.

Generative adversarial training for imitation learning has shown promising results on high-dimensional and continuous control tasks. This paradigm is based on reducing the imitation learning problem to the density matching problem, where the agent iteratively refines the policy to match the empirical state-action visitation frequency of the expert demonstration. Although this approach has shown to robustly learn to imitate even with scarce demonstration, one must still address the inherent challenge that collecting trajectory samples in each iteration is a costly operation. To address this issue, we first propose a Bayesian formulation of generative adversarial imitation learning (GAIL), where the imitation policy and the cost function are represented as stochastic neural networks. Then, we show that we can significantly enhance the sample efficiency of GAIL leveraging the predictive density of the cost, on an extensive set of imitation learning tasks with high-dimensional states and actions.

Imitation learning algorithms can be used to learn a policy from expert demonstrations without access to a reward signal.

Cinematic camera control demands a balance of precision and artistry - qualities that are difficult to encode through handcrafted reward functions. While reinforcement learning (RL) has been applied to robotic filmmaking, its reliance on bespoke rewards and extensive tuning limits creative usability. We propose a Learning from Demonstration (LfD) approach using Generative Adversarial Imitation Learning (GAIL) to automate dolly-in shots with a free-roaming, ground-based filming robot. Expert trajectories are collected via joystick teleoperation in simulation, capturing smooth, expressive motion without explicit objective design. Trained exclusively on these demonstrations, our GAIL policy outperforms a PPO baseline in simulation, achieving higher rewards, faster convergence, and lower variance. Crucially, it transfers directly to a real-world robot without fine-tuning, achieving more consistent framing and subject alignment than a prior TD3-based method. These results show that LfD offers a robust, reward-free alternative to RL in cinematic domains, enabling real-time deployment with minimal technical effort. Our pipeline brings intuitive, stylized camera control within reach of creative professionals, bridging the gap between artistic intent and robotic autonomy.

Reinforcement learning for multi-goal robot manipulation tasks poses significant challenges due to the diversity and complexity of the goal space. Techniques such as Hindsight Experience Replay (HER) have been introduced to improve learning efficiency for such tasks. More recently, researchers have combined HER with advanced imitation learning methods such as Generative Adversarial Imitation Learning (GAIL) to integrate demonstration data and accelerate training speed. However, demonstration data often fails to provide enough coverage for the goal space, especially when acquired from human teleoperation. This biases the learning-from-demonstration process toward mastering easier sub-tasks instead of tackling the more challenging ones. In this work, we present Goal-based Self-Adaptive Generative Adversarial Imitation Learning (Goal-SAGAIL), a novel framework specifically designed for multi-goal robot manipulation tasks. By integrating self-adaptive learning principles with goal-conditioned GAIL, our approach enhances imitation learning efficiency, even when limited, suboptimal demonstrations are available. Experimental results validate that our method significantly improves learning efficiency across various multi-goal manipulation scenarios -- including complex in-hand manipulation tasks -- using suboptimal demonstrations provided by both simulation and human experts.

Generative Adversarial Imitation Learning (GAIL) provides a promising approach to training a generative policy to imitate a demonstrator. It uses on-policy Reinforcement Learning (RL) to optimize a reward signal derived from an adversarial discriminator. However, optimizing GAIL is difficult in practise, with the training loss oscillating during training, slowing convergence. This optimization instability can prevent GAIL from finding a good policy, harming its final performance. In this paper, we study GAIL's optimization from a control-theoretic perspective.

Additional Feedback: My other main concern is that the objective in Eq. (5) is badly motivated and the implications are under underexplored. The imitation learning objective is notoriously ill-defined and a large part of the literature focuses on introducing objectives that produce good behavior. The notion of finding the "best" f-divergence therefore requires us to state what we are optimizing for, which the authors don't do very explicitly. On line 38, the authors mention that an imitation learning method which uses a fixed divergence method is likely to learn a sub-optimal policy, but the notion of optimality does not exist without a given divergence. For example, whether mode-seeking or mode-covering behavior is better is entirely dependent on context that the agent does not have. Either solution could be better.

After reading the authors' rebuttal, the reviewers discussed their concerns about this paper. Ultimately, a consensus was not reached as reviewer #3 feels that some of her/his concerns were not properly addressed in the authors' feedback. The other reviewers are positive with respect to the paper (especially thanks to the promising experimental results), but they share one of the concerns of reviewer #3, i.e., the definition of optimal f-divergence'' and the convergence properties of the proposed approach. I agree with them that the paper has merits and the ideas contained in the paper are interesting, so I propose to accept it, but I recommend that the authors take the issues raised in the reviews seriously and address them carefully in the final version of the paper.