Domain randomization in reinforcement learning is an established technique for increasing the robustness of control policies trained in simulation. By randomizing environment properties during training, the learned policy can become robust to uncertainties along the randomized dimensions. While the environment distribution is typically specified by hand, in this paper we investigate automatically discovering a sampling distribution via entropy-regularized reward maximization of a normalizing-flow-based neural sampling distribution. We show that this architecture is more flexible and provides greater robustness than existing approaches that learn simpler, parameterized sampling distributions, as demonstrated in six simulated and one real-world robotics domain. Lastly, we explore how these learned sampling distributions, combined with a privileged value function, can be used for out-of-distribution detection in an uncertainty-aware multi-step manipulation planner.

Integrated task and motion planning (TAMP) has proven to be a valuable approach to generalizable long-horizon robotic manipulation and navigation problems. However, the typical TAMP problem formulation assumes full observability and deterministic action effects. These assumptions limit the ability of the planner to gather information and make decisions that are risk-aware. We propose a strategy for TAMP with Uncertainty and Risk Awareness (TAMPURA) that is capable of efficiently solving long-horizon planning problems with initial-state and action outcome uncertainty, including problems that require information gathering and avoiding undesirable and irreversible outcomes. Our planner reasons under uncertainty at both the abstract task level and continuous controller level. Given a set of closed-loop goal-conditioned controllers operating in the primitive action space and a description of their preconditions and potential capabilities, we learn a high-level abstraction that can be solved efficiently and then refined to continuous actions for execution. We demonstrate our approach on several robotics problems where uncertainty is a crucial factor and show that reasoning under uncertainty in these problems outperforms previously proposed determinized planning, direct search, and reinforcement learning strategies. Lastly, we demonstrate our planner on two real-world robotics problems using recent advancements in probabilistic perception.

Robots cannot yet match humans' ability to rapidly learn the shapes of novel 3D objects and recognize them robustly despite clutter and occlusion. We present Bayes3D, an uncertainty-aware perception system for structured 3D scenes, that reports accurate posterior uncertainty over 3D object shape, pose, and scene composition in the presence of clutter and occlusion. Bayes3D delivers these capabilities via a novel hierarchical Bayesian model for 3D scenes and a GPU-accelerated coarse-to-fine sequential Monte Carlo algorithm. Quantitative experiments show that Bayes3D can learn 3D models of novel objects from just a handful of views, recognizing them more robustly and with orders of magnitude less training data than neural baselines, and tracking 3D objects faster than real time on a single GPU. We also demonstrate that Bayes3D learns complex 3D object models and accurately infers 3D scene composition when used on a Panda robot in a tabletop scenario.

To facilitate the development of new models to bridge the gap between machine and human social intelligence, the recently proposed Baby Intuitions Benchmark (arXiv:2102.11938) provides a suite of tasks designed to evaluate commonsense reasoning about agents' goals and actions that even young infants exhibit. Here we present a principled Bayesian solution to this benchmark, based on a hierarchically Bayesian Theory of Mind (HBToM). By including hierarchical priors on agent goals and dispositions, inference over our HBToM model enables few-shot learning of the efficiency and preferences of an agent, which can then be used in commonsense plausibility judgements about subsequent agent behavior. This approach achieves near-perfect accuracy on most benchmark tasks, outperforming deep learning and imitation learning baselines while producing interpretable human-like inferences, demonstrating the advantages of structured Bayesian models of human social cognition.

We present 3DP3, a framework for inverse graphics that uses inference in a structured generative model of objects, scenes, and images. 3DP3 uses (i) voxel models to represent the 3D shape of objects, (ii) hierarchical scene graphs to decompose scenes into objects and the contacts between them, and (iii) depth image likelihoods based on real-time graphics. Given an observed RGB-D image, 3DP3's inference algorithm infers the underlying latent 3D scene, including the object poses and a parsimonious joint parametrization of these poses, using fast bottom-up pose proposals, novel involutive MCMC updates of the scene graph structure, and, optionally, neural object detectors and pose estimators. We show that 3DP3 enables scene understanding that is aware of 3D shape, occlusion, and contact structure. Our results demonstrate that 3DP3 is more accurate at 6DoF object pose estimation from real images than deep learning baselines and shows better generalization to challenging scenes with novel viewpoints, contact, and partial observability.

Probabilistic graphical models (PGMs) provide a compact representation of knowledge that can be queried in a flexible way: after learning the parameters of a graphical model, new probabilistic queries can be answered at test time without retraining. However, learning undirected graphical models is notoriously hard due to the intractability of the partition function. For directed models, a popular approach is to use variational autoencoders, but there is no systematic way to choose the encoder architecture given the PGM, and the encoder only amortizes inference for a single probabilistic query (i.e., new queries require separate training). We introduce Query Training (QT), a systematic method to turn any PGM structure (directed or not, with or without hidden variables) into a trainable inference network. This single network can approximate any inference query. We demonstrate experimentally that QT can be used to learn a challenging 8-connected grid Markov random field with hidden variables and that it consistently outperforms the state-of-the-art AdVIL when tested on three undirected models across multiple datasets.

For an intelligent agent to flexibly and efficiently operate in complex environments, they must be able to reason at multiple levels of temporal, spatial, and conceptual abstraction. At the lower levels, the agent must interpret their proprioceptive inputs and control their muscles, and at the higher levels, the agent must select goals and plan how they will achieve those goals. It is clear that each of these types of reasoning is amenable to different types of representations, algorithms, and inputs. In this work, we introduce a simple, three-level hierarchical architecture that reflects these distinctions. The low-level controller operates on the continuous proprioceptive inputs, using model-free learning to acquire useful behaviors. These in turn induce a set of mid-level dynamics, which are learned by the mid-level controller and used for model-predictive control, to select a behavior to activate at each timestep. The high-level controller leverages a discrete, graph representation for goal selection and path planning to specify targets for the mid-level controller. We apply our method to a series of navigation tasks in the Mujoco Ant environment, consistently demonstrating significant improvements in sample-efficiency compared to prior model-free, model-based, and hierarchical RL methods. Finally, as an illustrative example of the advantages of our architecture, we apply our method to a complex maze environment that requires efficient exploration and long-horizon planning.

Sequence modeling is a fundamental real-world problem with a wide range of applications. Recurrent neural networks (RNNs) are currently preferred in sequence prediction tasks due to their ability to model long-term and variable order dependencies. However, RNNs have disadvantages in several applications because of their inability to natively handle uncertainty, and because of the inscrutable internal representations. Probabilistic sequence models like Hidden Markov Models (HMM) have the advantage of more interpretable representations and the ability to handle uncertainty. Although overcomplete HMMs with many more hidden states compared to the observed states can, in theory, model long-term temporal dependencies [23], training HMMs is challenging due to credit diffusion [3]. For this reason, simpler and inflexible n-gram models are preferred to HMMs for tasks like language modeling. Tensor decomposition methods [1] have been suggested for the learning of HMMs in order to overcome the credit diffusion problem, but current methods are not applicable to the overcomplete setting where the full-rank requirements on the transition and emission matrices are not fulfilled. Recently there has been renewed interest in the topic of training overcomplete HMMs for higher-order dependencies with the expectation that sparsity structures could potentially alleviate the credit diffusion problem [23]. In this paper we demonstrate that a particular sparsity structure on the emission matrix can help HMMs learn higher-order temporal structure using the standard Expectation-Maximization algorithms [26] (Baum-Welch) and its online variants.