While problem solving is a standard evaluation task for foundation models, a crucial component of problem solving -- actively and strategically gathering information to test hypotheses -- has not been closely investigated. To assess the information gathering abilities of foundation models in interactive environments, we introduce a framework in which a model must determine the factors influencing a hidden reward function by iteratively reasoning about its previously gathered information and proposing its next exploratory action to maximize information gain at each step. We implement this framework in both a text-based environment, which offers a tightly controlled setting and enables high-throughput parameter sweeps, and in an embodied 3D environment, which requires addressing complexities of multi-modal interaction more relevant to real-world applications. We further investigate whether approaches such as self-correction and increased inference time improve information gathering efficiency. In a relatively simple task that requires identifying a single rewarding feature, we find that LLM's information gathering capability is close to optimal. However, when the model must identify a conjunction of rewarding features, performance is suboptimal. The hit in performance is due partly to the model translating task description to a policy and partly to the model's effectiveness in using its in-context memory. Performance is comparable in both text and 3D embodied environments, although imperfect visual object recognition reduces its accuracy in drawing conclusions from gathered information in the 3D embodied case. For single-feature-based rewards, we find that smaller models curiously perform better; for conjunction-based rewards, incorporating self correction into the model improves performance.

Building embodied AI systems that can follow arbitrary language instructions in any 3D environment is a key challenge for creating general AI. Accomplishing this goal requires learning to ground language in perception and embodied actions, in order to accomplish complex tasks. The Scalable, Instructable, Multiworld Agent (SIMA) project tackles this by training agents to follow free-form instructions across a diverse range of virtual 3D environments, including curated research environments as well as open-ended, commercial video games. Our goal is to develop an instructable agent that can accomplish anything a human can do in any simulated 3D environment. Our approach focuses on language-driven generality while imposing minimal assumptions. Our agents interact with environments in real-time using a generic, human-like interface: the inputs are image observations and language instructions and the outputs are keyboard-and-mouse actions. This general approach is challenging, but it allows agents to ground language across many visually complex and semantically rich environments while also allowing us to readily run agents in new environments. In this paper we describe our motivation and goal, the initial progress we have made, and promising preliminary results on several diverse research environments and a variety of commercial video games.

Abstract--Quadruped locomotion is rapidly maturing to a degree where robots are able to realise highly dynamic manoeuvres. However, current planners are unable to vary key gait parameters of the in-swing feet midair. In this work we address this limitation and show that it is pivotal in increasing controller robustness by learning a latent space capturing the key stance phases constituting a particular gait. This is achieved via a generative model trained on a single trot style, which encourages disentanglement such that application of a drive signal to a single dimension of the latent state induces holistic plans synthesising a continuous variety of trot styles. We demonstrate that specific properties of the drive signal map directly to gait parameters such as cadence, footstep height and full stance duration. Due to the nature of our approach these synthesised gaits are continuously variable online during robot operation. The use of a generative model facilitates the detection and mitigation of disturbances to provide a versatile and robust planning framework. We evaluate our approach on two versions of the real ANYmal quadruped robots and demonstrate that our method achieves a continuous blend of dynamic trot styles whilst being robust and reactive to external perturbations. Figure 1: Using a variational auto-encoder (VAE), our approach learns a structured latent space capturing key stance phases constituting a particular gait. The space is disentangled to I. I This approach by advances in optimisation-based [1]-[5] and reinforcement allows for precise base twist control and readily transfers from learning-based methods [6]-[8], quadrupeds are now able to ANYmal B to ANYmal C, a dynamically dissimilar robot, robustly plan and perform dynamic manoeuvres, making them without retraining. Additionally, we measure disturbances as an increasingly popular and reliable choice for tasks such out of distribution seen during training and adjust cadence as as inspection, monitoring, search and rescue or goods delivery a rudimentary, but effective response. However, despite recent advances, important limitations remain. Due to the complexity of the system, models used for gait planning and control are often overly contact schedules [1], [9]. Mathieu Geisert is with Agility Robotics, U.S.A. Work done while at Martin Engelcke is with DeepMind Technologies Ltd., London, U.K. Work done while at Oxford. Personal use of this material is permitted. These are often characterise and react to external perturbations. A large impulse computationally expensive [3], [4] meaning that varying the applied to the robot's base triggers a spike in the gait parameters is not achievable in real time. A limitation Evidence Lower Bound (ELBO) which clearly identifies the of all these methods is that they are unable to adjust key disturbance as out of the distribution seen during training.

Modelling functions of sets, or equivalently, permutation-invariant functions, is a long-standing challenge in machine learning. Deep Sets is a popular method which is known to be a universal approximator for continuous set functions. We provide a theoretical analysis of Deep Sets which shows that this universal approximation property is only guaranteed if the model's latent space is sufficiently high-dimensional. If the latent space is even one dimension lower than necessary, there exist piecewise-affine functions for which Deep Sets performs no better than a na\"ive constant baseline, as judged by worst-case error. Deep Sets may be viewed as the most efficient incarnation of the Janossy pooling paradigm. We identify this paradigm as encompassing most currently popular set-learning methods. Based on this connection, we discuss the implications of our results for set learning more broadly, and identify some open questions on the universality of Janossy pooling in general.

Advances in object-centric generative models (OCGMs) have culminated in the development of a broad range of methods for unsupervised object segmentation and interpretable object-centric scene generation. These methods, however, are limited to simulated and real-world datasets with limited visual complexity. Moreover, object representations are often inferred using RNNs which do not scale well to large images or iterative refinement which avoids imposing an unnatural ordering on objects in an image but requires the a priori initialisation of a fixed number of object representations. In contrast to established paradigms, this work proposes an embedding-based approach in which embeddings of pixels are clustered in a differentiable fashion using a stochastic, non-parametric stick-breaking process. Similar to iterative refinement, this clustering procedure also leads to randomly ordered object representations, but without the need of initialising a fixed number of clusters a priori. This is used to develop a new model, GENESIS-V2, which can infer a variable number of object representations without using RNNs or iterative refinement. We show that GENESIS-V2 outperforms previous methods for unsupervised image segmentation and object-centric scene generation on established synthetic datasets as well as more complex real-world datasets.

A range of methods with suitable inductive biases exist to learn interpretable object-centric representations of images without supervision. However, these are largely restricted to visually simple images; robust object discovery in real-world sensory datasets remains elusive. To increase the understanding of such inductive biases, we empirically investigate the role of "reconstruction bottlenecks" for scene decomposition in GENESIS, a recent VAE-based model. We show such bottlenecks determine reconstruction and segmentation quality and critically influence model behaviour.

Generative models are emerging as promising tools in robotics and reinforcement learning. Yet, even though tasks in these domains typically involve distinct objects, most state-of-the-art methods do not explicitly capture the compositional nature of visual scenes. Two exceptions, MONet and IODINE, decompose scenes into objects in an unsupervised fashion via a set of latent variables. Their underlying generative processes, however, do not account for component interactions. Hence, neither of them allows for principled sampling of coherent scenes. Here we present GENESIS, the first object-centric generative model of visual scenes capable of both decomposing and generating complete scenes by explicitly capturing relationships between scene components. GENESIS parameterises a spatial GMM over pixels which is encoded by component-wise latent variables that are inferred sequentially or sampled from an autoregressive prior. We train GENESIS on two publicly available datasets and probe the information in the latent representations through a set of classification tasks, outperforming several baselines.

Recent work on the representation of functions on sets has considered the use of summation in a latent space to enforce permutation invariance. In particular, it has been conjectured that the dimension of this latent space may remain fixed as the cardinality of the sets under consideration increases. However, we demonstrate that the analysis leading to this conjecture requires mappings which are highly discontinuous and argue that this is only of limited practical use. Motivated by this observation, we prove that an implementation of this model via continuous mappings (as provided by e.g. neural networks or Gaussian processes) actually imposes a constraint on the dimensionality of the latent space. Practical universal function representation for set inputs can only be achieved with a latent dimension at least the size of the maximum number of input elements.