Existing off-policy reinforcement learning algorithms typically necessitate an explicit state-action-value function representation, which becomes problematic in high-dimensional action spaces. These algorithms often encounter challenges where they struggle with the curse of dimensionality, as maintaining a state-action-value function in such spaces becomes data-inefficient. In this work, we propose a novel off-policy trust region optimization approach, called Vlearn, that eliminates the requirement for an explicit state-action-value function. Instead, we demonstrate how to efficiently leverage just a state-value function as the critic, thus overcoming several limitations of existing methods. By doing so, Vlearn addresses the computational challenges posed by high-dimensional action spaces. Furthermore, Vlearn introduces an efficient approach to address the challenges associated with pure state-value function learning in the off-policy setting. This approach not only simplifies the implementation of off-policy policy gradient algorithms but also leads to consistent and robust performance across various benchmark tasks. Specifically, by removing the need for a state-action-value function Vlearn simplifies the learning process and allows for more efficient exploration and exploitation in complex environments
Stochastic gradient-based optimization is crucial to optimize neural networks. While popular approaches heuristically adapt the step size and direction by rescaling gradients, a more principled approach to improve optimizers requires second-order information. Such methods precondition the gradient using the objective's Hessian. Yet, computing the Hessian is usually expensive and effectively using second-order information in the stochastic gradient setting is non-trivial. We propose using Information-Theoretic Trust Region Optimization (arTuRO) for improved updates with uncertain second-order information. By modeling the network parameters as a Gaussian distribution and using a Kullback-Leibler divergence-based trust region, our approach takes bounded steps accounting for the objective's curvature and uncertainty in the parameters. Before each update, it solves the trust region problem for an optimal step size, resulting in a more stable and faster optimization process. We approximate the diagonal elements of the Hessian from stochastic gradients using a simple recursive least squares approach, constructing a model of the expected Hessian over time using only first-order information. We show that arTuRO combines the fast convergence of adaptive moment-based optimization with the generalization capabilities of SGD.
Bayesian deep learning (BDL) is a promising approach to achieve well-calibrated predictions on distribution-shifted data. Nevertheless, there exists no large-scale survey that evaluates recent SOTA methods on diverse, realistic, and challenging benchmark tasks in a systematic manner. To provide a clear picture of the current state of BDL research, we evaluate modern BDL algorithms on real-world datasets from the WILDS collection containing challenging classification and regression tasks, with a focus on generalization capability and calibration under distribution shift. We compare the algorithms on a wide range of large, convolutional and transformer-based neural network architectures. In particular, we investigate a signed version of the expected calibration error that reveals whether the methods are over- or under-confident, providing further insight into the behavior of the methods. Further, we provide the first systematic evaluation of BDL for fine-tuning large pre-trained models, where training from scratch is prohibitively expensive. Finally, given the recent success of Deep Ensembles, we extend popular single-mode posterior approximations to multiple modes by the use of ensembles. While we find that ensembling single-mode approximations generally improves the generalization capability and calibration of the models by a significant margin, we also identify a failure mode of ensembles when finetuning large transformer-based language models. In this setting, variational inference based approaches such as last-layer Bayes By Backprop outperform other methods in terms of accuracy by a large margin, while modern approximate inference algorithms such as SWAG achieve the best calibration.
Bayesian deep learning (BDL) is a promising approach to achieve well-calibrated predictions on distribution-shifted data. Nevertheless, there exists no large-scale survey that evaluates recent SOTA methods on diverse, realistic, and challenging benchmark tasks in a systematic manner. To provide a clear picture of the current state of BDL research, we evaluate modern BDL algorithms on real-world datasets from the WILDS collection containing challenging classification and regression tasks, with a focus on generalization capability and calibration under distribution shift. We compare the algorithms on a wide range of large, convolutional and transformer-based neural network architectures. In particular, we investigate a signed version of the expected calibration error that reveals whether the methods are over- or under-confident, providing further insight into the behavior of the methods. Further, we provide the first systematic evaluation of BDL for fine-tuning large pre-trained models, where training from scratch is prohibitively expensive. Finally, given the recent success of Deep Ensembles, we extend popular single-mode posterior approximations to multiple modes by the use of ensembles. While we find that ensembling single-mode approximations generally improves the generalization capability and calibration of the models by a significant margin, we also identify a failure mode of ensembles when finetuning large transformer-based language models. In this setting, variational inference based approaches such as last-layer Bayes By Backprop outperform other methods in terms of accuracy by a large margin, while modern approximate inference algorithms such as SWAG achieve the best calibration.
Learning skills by imitation is a promising concept for the intuitive teaching of robots. A common way to learn such skills is to learn a parametric model by maximizing the likelihood given the demonstrations. Yet, human demonstrations are often multi-modal, i.e., the same task is solved in multiple ways which is a major challenge for most imitation learning methods that are based on such a maximum likelihood (ML) objective. The ML objective forces the model to cover all data, it prevents specialization in the context space and can cause mode-averaging in the behavior space, leading to suboptimal or potentially catastrophic behavior. Here, we alleviate those issues by introducing a curriculum using a weight for each data point, allowing the model to specialize on data it can represent while incentivizing it to cover as much data as possible by an entropy bonus. We extend our algorithm to a Mixture of (linear) Experts (MoE) such that the single components can specialize on local context regions, while the MoE covers all data points. We evaluate our approach in complex simulated and real robot control tasks and show it learns from versatile human demonstrations and significantly outperforms current SOTA methods. A reference implementation can be found at https://github.com/intuitive-robots/ml-cur
In many scenarios, observations from more than one sensor modality are available for reinforcement learning (RL). For example, many agents can perceive their internal state via proprioceptive sensors but must infer the environment's state from high-dimensional observations such as images. For image-based RL, a variety of self-supervised representation learning approaches exist to improve performance and sample complexity. These approaches learn the image representation in isolation. However, including proprioception can help representation learning algorithms to focus on relevant aspects and guide them toward finding better representations. Hence, in this work, we propose using Recurrent State Space Models to fuse all available sensory information into a single consistent representation. We combine reconstruction-based and contrastive approaches for training, which allows using the most appropriate method for each sensor modality. For example, we can use reconstruction for proprioception and a contrastive loss for images. We demonstrate the benefits of utilizing proprioception in learning representations for RL on a large set of experiments. Furthermore, we show that our joint representations significantly improve performance compared to a post hoc combination of image representations and proprioception.
Improved state space models, such as Recurrent State Space Models (RSSMs), are a key factor behind recent advances in model-based reinforcement learning (RL). Yet, despite their empirical success, many of the underlying design choices are not well understood. We show that RSSMs use a suboptimal inference scheme and that models trained using this inference overestimate the aleatoric uncertainty of the ground truth system. We find this overestimation implicitly regularizes RSSMs and allows them to succeed in model-based RL. We postulate that this implicit regularization fulfills the same functionality as explicitly modeling epistemic uncertainty, which is crucial for many other model-based RL approaches. Yet, overestimating aleatoric uncertainty can also impair performance in cases where accurately estimating it matters, e.g., when we have to deal with occlusions, missing observations, or fusing sensor modalities at different frequencies. Moreover, the implicit regularization is a side-effect of the inference scheme and not the result of a rigorous, principled formulation, which renders analyzing or improving RSSMs difficult. Thus, we propose an alternative approach building on well-understood components for modeling aleatoric and epistemic uncertainty, dubbed Variational Recurrent Kalman Network (VRKN). This approach uses Kalman updates for exact smoothing inference in a latent space and Monte Carlo Dropout to model epistemic uncertainty. Due to the Kalman updates, the VRKN can naturally handle missing observations or sensor fusion problems with varying numbers of observations per time step. Our experiments show that using the VRKN instead of the RSSM improves performance in tasks where appropriately capturing aleatoric uncertainty is crucial while matching it in the deterministic standard benchmarks.
Humans intuitively solve tasks in versatile ways, varying their behavior in terms of trajectory-based planning and for individual steps. Thus, they can easily generalize and adapt to new and changing environments. Current Imitation Learning algorithms often only consider unimodal expert demonstrations and act in a state-action-based setting, making it difficult for them to imitate human behavior in case of versatile demonstrations. Instead, we combine a mixture of movement primitives with a distribution matching objective to learn versatile behaviors that match the expert's behavior and versatility. To facilitate generalization to novel task configurations, we do not directly match the agent's and expert's trajectory distributions but rather work with concise geometric descriptors which generalize well to unseen task configurations. We empirically validate our method on various robot tasks using versatile human demonstrations and compare to imitation learning algorithms in a state-action setting as well as a trajectory-based setting. We find that the geometric descriptors greatly help in generalizing to new task configurations and that combining them with our distribution-matching objective is crucial for representing and reproducing versatile behavior.
Recurrent State-space models (RSSMs) are highly expressive models for learning patterns in time series data and system identification. However, these models assume that the dynamics are fixed and unchanging, which is rarely the case in real-world scenarios. Many control applications often exhibit tasks with similar but not identical dynamics which can be modeled as a latent variable. We introduce the Hidden Parameter Recurrent State Space Models (HiP-RSSMs), a framework that parametrizes a family of related dynamical systems with a low-dimensional set of latent factors. We present a simple and effective way of learning and performing inference over this Gaussian graphical model that avoids approximations like variational inference. We show that HiP-RSSMs outperforms RSSMs and competing multi-task models on several challenging robotic benchmarks both on real-world systems and simulations.
It is well-known that inverse dynamics models can improve tracking performance in robot control. These models need to precisely capture the robot dynamics, which consist of well-understood components, e.g., rigid body dynamics, and effects that remain challenging to capture, e.g., stick-slip friction and mechanical flexibilities. Such effects exhibit hysteresis and partial observability, rendering them, particularly challenging to model. Hence, hybrid models, which combine a physical prior with data-driven approaches are especially well-suited in this setting. We present a novel hybrid model formulation that enables us to identify fully physically consistent inertial parameters of a rigid body dynamics model which is paired with a recurrent neural network architecture, allowing us to capture unmodeled partially observable effects using the network memory. We compare our approach against state-of-the-art inverse dynamics models on a 7 degree of freedom manipulator. Using data sets obtained through an optimal experiment design approach, we study the accuracy of offline torque prediction and generalization capabilities of joint learning methods. In control experiments on the real system, we evaluate the model as a feed-forward term for impedance control and show the feedback gains can be drastically reduced to achieve a given tracking accuracy.