NeurIPS 2024 Papers

Data augmentation creates new data points by transforming the original ones for an reinforcement learning (RL) agent to learn from, which has been shown to be effective for the objective of improving data efficiency of RL for continuous control. Prior work towards this objective has been largely restricted to perturbation-based data augmentation where new data points are created by perturbing the original ones, which has been impressively effective for tasks where the RL agent observe control states as images with perturbations including random cropping, shifting, etc. This work focuses on state-based control, where the RL agent can directly observe raw kinematic and task features, and considers an alternative data augmentation applied to these features based on Euclidean symmetries under transformations like rotations. We show that the default state features used in exiting benchmark tasks that are based on joint configurations are not amenable to Euclidean transformations. We therefore advocate using state features based on configurations of the limbs (i.e., rigid bodies connected by joints) that instead provides rich augmented data under Euclidean transformations. With minimal hyperparameter tuning, we show this new Euclidean data augmentation strategy significantly improve both data efficiency and asymptotic performance of RL on a wide range of continuous control tasks.

Conditional independence tests are crucial across various disciplines in determining the independence of an outcome variable $Y$ from a treatment variable $X$, conditioning on a set of confounders $Z$. The Conditional Randomization Test (CRT) offers a powerful framework for such testing by assuming known distributions of $X \mid Z$; it controls the Type-I error exactly, allowing for the use of flexible, black-box test statistics. In practice, testing for conditional independence often involves using data from a source population to draw conclusions about a target population. This can be challenging due to covariate shift---differences in the distribution of $X$, $Z$, and surrogate variables, which can affect the conditional distribution of $Y \mid X, Z$---rendering traditional CRT approaches invalid. To address this issue, we propose a novel Covariate Shift Corrected Pearson Chi-squared Conditional Randomization (csPCR) test. This test adapts to covariate shifts by integrating importance weights and employing the control variates method to reduce variance in the test statistics and thus enhance power. Theoretically, we establish that the csPCR test controls the Type-I error asymptotically. Empirically, through simulation studies, we demonstrate that our method not only maintains control over Type-I errors but also exhibits superior power, confirming its efficacy and practical utility in real-world scenarios where covariate shifts are prevalent. Finally, we apply our methodology to a real-world dataset to assess the impact of a COVID-19 treatment on the 90-day mortality rate among patients.

Generalizing across robot embodiments and tasks is crucial for adaptive robotic systems. Modular policy learning approaches adapt to new embodiments but are limited to specific tasks, while few-shot imitation learning (IL) approaches often focus on a single embodiment. In this paper, we introduce a few-shot behavior cloning framework to simultaneously generalize to unseen embodiments and tasks using a few (e.g., five) reward-free demonstrations. Our framework leverages a joint-level input-output representation to unify the state and action spaces of heterogeneous embodiments and employs a novel structure-motion state encoder that is parameterized to capture both shared knowledge across all embodiments and embodiment-specific knowledge. A matching-based policy network then predicts actions from a few demonstrations, producing an adaptive policy that is robust to over-fitting. Evaluated in the DeepMind Control suite, our framework termed Meta-Controller demonstrates superior few-shot generalization to unseen embodiments and tasks over modular policy learning and few-shot IL approaches.

Reinforcement learning (RL) demonstrates superior potential over traditional flight control methods for fixed-wing aircraft, particularly under extreme operational conditions. However, the high demand for training samples and the lack of efficient computation in existing simulators hinder its further application. In this paper, we introduce NeuralPlane, the first benchmark platform for large-scale parallel simulations of fixed-wing aircraft. NeuralPlane significantly boosts high-fidelity simulation via GPU-accelerated Flight Dynamics Model (FDM) computation, achieving a single-step simulation time of just 0.2 seconds at a parallel scale of $10^{6}$, far exceeding current platforms. We also provide clear code templates, comprehensive evaluation/visualization tools and hierarchical frameworks for integrating RL and traditional control methods. We believe that NeuralPlane can accelerate the development of RL-based fixed-wing flight control and serve as a new challenging benchmark for the RL community. Our NeuralPlane is open-source and accessible at https://github.com/xuecy22/NeuralPlane.

This paper introduces the risk-sensitive control as inference (RCaI) that extends CaI by using Rényi divergence variational inference. RCaI is shown to be equivalent to log-probability regularized risk-sensitive control, which is an extension of the maximum entropy (MaxEnt) control. We also prove that the risk-sensitive optimal policy can be obtained by solving a soft Bellman equation, which reveals several equivalences between RCaI, MaxEnt control, the optimal posterior for CaI, and linearly-solvable control. Moreover, based on RCaI, we derive the risk-sensitive reinforcement learning (RL) methods: the policy gradient and the soft actor-critic. As the risk-sensitivity parameter vanishes, we recover the risk-neutral CaI and RL, which means that RCaI is a unifying framework. Furthermore, we give another risk-sensitive generalization of the MaxEnt control using Rényi entropy regularization. We show that in both of our extensions, the optimal policies have the same structure even though the derivations are very different.

We present NeuralFluid, a novel framework to explore neural control and design of complex fluidic systems with dynamic solid boundaries. Our system features a fast differentiable Navier-Stokes solver with solid-fluid interface handling, a low-dimensional differentiable parametric geometry representation, a control-shape co-design algorithm, and gym-like simulation environments to facilitate various fluidic control design applications. Additionally, we present a benchmark of design, control, and learning tasks on high-fidelity, high-resolution dynamic fluid environments that pose challenges for existing differentiable fluid simulators. These tasks include designing the control of artificial hearts, identifying robotic end-effector shapes, and controlling a fluid gate. By seamlessly incorporating our differentiable fluid simulator into a learning framework, we demonstrate successful design, control, and learning results that surpass gradient-free solutions in these benchmark tasks.

Cinematographers adeptly capture the essence of the world, crafting compelling visual narratives through intricate camera movements. Witnessing the strides made by large language models in perceiving and interacting with the 3D world, this study explores their capability to control cameras with human language guidance. We introduce ChatCam, a system that navigates camera movements through conversations with users, mimicking a professional cinematographer's workflow. To achieve this, we propose CineGPT, a GPT-based autoregressive model for text-conditioned camera trajectory generation. We also develop an Anchor Determinator to ensure precise camera trajectory placement. ChatCam understands user requests and employs our proposed tools to generate trajectories, which can be used to render high-quality video footage on radiance field representations. Our experiments, including comparisons to state-of-the-art approaches and user studies, demonstrate our approach's ability to interpret and execute complex instructions for camera operation, showing promising applications in real-world production settings. Project page: https://xinhangliu.com/chatcam.

Dimension reduction (DR) is an important and widely studied technique in exploratory data analysis. However, traditional DR methods are not applicable to datasets with with a contrastive structure, where data are split into a foreground group of interest (case or treatment group), and a background group (control group). This type of data, common in biomedical studies, necessitates contrastive dimension reduction (CDR) methods to effectively capture information unique to or enriched in the foreground group relative to the background group. Despite the development of various CDR methods, two critical questions remain underexplored: when should these methods be applied, and how can the information unique to the foreground group be quantified? In this work, we address these gaps by proposing a hypothesis test to determine the existence of contrastive information, and introducing a contrastive dimension estimator (CDE) to quantify the unique components in the foreground group. We provide theoretical support for our methods and validate their effectiveness through extensive simulated, semi-simulated, and real experiments involving images, gene expressions, protein expressions, and medical sensors, demonstrating their ability to identify the unique information in the foreground group.

We cast real-world humanoid control as a next token prediction problem, akin to predicting the next word in language. Our model is a causal transformer trained via autoregressive prediction of sensorimotor sequences. To account for the multi-modal nature of the data, we perform prediction in a modality-aligned way, and for each input token predict the next token from the same modality. This general formulation enables us to leverage data with missing modalities, such as videos without actions. We train our model on a dataset of sequences from a prior neural network policy, a model-based controller, motion capture, and YouTube videos of humans. We show that our model enables a real humanoid robot to walk in San Francisco zero-shot. Our model can transfer to the real world even when trained on only 27 hours of walking data, and can generalize to commands not seen during training. These findings suggest a promising path toward learning challenging real-world control tasks by generative modeling of sensorimotor sequences.

We present a method for controlling a simulated humanoid to grasp an object and move it to follow an object's trajectory. Due to the challenges in controlling a humanoid with dexterous hands, prior methods often use a disembodied hand and only consider vertical lifts or short trajectories. This limited scope hampers their applicability for object manipulation required for animation and simulation. To close this gap, we learn a controller that can pick up a large number (>1200) of objects and carry them to follow randomly generated trajectories. Our key insight is to leverage a humanoid motion representation that provides human-like motor skills and significantly speeds up training. Using only simplistic reward, state, and object representations, our method shows favorable scalability on diverse objects and trajectories. For training, we do not need a dataset of paired full-body motion and object trajectories. At test time, we only require the object mesh and desired trajectories for grasping and transporting. To demonstrate the capabilities of our method, we show state-of-the-art success rates in following object trajectories and generalizing to unseen objects. Code and models will be released.

Animals survive in dynamic environments changing at arbitrary timescales, but such data distribution shifts are a challenge to neural networks. To adapt to change, neural systems may change a large number of parameters, which is a slow process involving forgetting past information. In contrast, animals leverage distribution changes to segment their stream of experience into tasks and associate them with internal task abstracts. Animals can then respond flexibly by selecting the appropriate task abstraction. However, how such flexible task abstractions may arise in neural systems remains unknown. Here, we analyze a linear gated network where the weights and gates are jointly optimized via gradient descent, but with neuron-like constraints on the gates including a faster timescale, non-negativity, and bounded activity. We observe that the weights self-organize into modules specialized for tasks or sub-tasks encountered, while the gates layer forms unique representations that switch the appropriate weight modules (task abstractions). We analytically reduce the learning dynamics to an effective eigenspace, revealing a virtuous cycle: fast adapting gates drive weight specialization by protecting previous knowledge, while weight specialization in turn increases the update rate of the gating layer. Task switching in the gating layer accelerates as a function of curriculum block size and task training, mirroring key findings in cognitive neuroscience. We show that the discovered task abstractions support generalization through both task and subtask composition, and we extend our findings to a non-linear network switching between two tasks. Overall, our work offers a theory of cognitive flexibility in animals as arising from joint gradient descent on synaptic and neural gating in a neural network architecture.

AI model alignment is crucial due to inadvertent biases in training data and the underspecified machine learning pipeline, where models with excellent test metrics may not meet end-user requirements. While post-training alignment via human feedback shows promise, these methods are often limited to generative AI settings where humans can interpret and provide feedback on model outputs. In traditional non-generative settings with numerical or categorical outputs, detecting misalignment through single-sample outputs remains challenging, and enforcing alignment during training requires repeating costly training processes. In this paper we consider an alternative strategy. We propose interpreting model alignment through property testing, defining an aligned model $f$ as one belonging to a subset $\mathcal{P}$ of functions that exhibit specific desired behaviors. We focus on post-processing a pre-trained model $f$ to better align with $\mathcal{P}$ using conformal risk control. Specifically, we develop a general procedure for converting queries for testing a given property $\mathcal{P}$ to a collection of loss functions suitable for use in a conformal risk control algorithm. We prove a probabilistic guarantee that the resulting conformal interval around $f$ contains a function approximately satisfying $\mathcal{P}$. We exhibit applications of our methodology on a collection of supervised learning datasets for (shape-constrained) properties such as monotonicity and concavity. The general procedure is flexible and can be applied to a wide range of desired properties. Finally, we prove that pre-trained models will always require alignment techniques even as model sizes or training data increase, as long as the training data contains even small biases.

We introduce the framework of performative control, where the policy chosen by the controller affects the underlying dynamics of the control system. This results in a sequence of policy-dependent system state data with policy-dependent temporal correlations. Following the recent literature on performative prediction \cite{perdomo2020performative}, we introduce the concept of a performatively stable control (PSC) solution. We first propose a sufficient condition for the performative control problem to admit a unique PSC solution with a problem-specific structure of distributional sensitivity propagation and aggregation. We further analyze the impacts of system stability on the existence of the PSC solution. Specifically, for {almost surely strongly stable} policy-dependent dynamics, the PSC solution exists if the sum of the distributional sensitivities is small enough. However, for almost surely unstable policy-dependent dynamics, the existence of the PSC solution will necessitate a temporally backward decaying of the distributional sensitivities. We finally provide a repeated stochastic gradient descent scheme that converges to the PSC solution and analyze its non-asymptotic convergence rate. Numerical results validate our theoretical analysis.

Neural Episodic Control is a powerful reinforcement learning framework that employs a differentiable dictionary to store non-parametric memories. It was inspired by episodic memory on the functional level, but lacks a direct theoretical connection to the associative memory models generally used to implement such a memory. We first show that the dictionary is an instance of the recently proposed Universal Hopfield Network framework. We then introduce a continuous approximation of the dictionary readout operation in order to derive two energy functions that are Lyapunov functions of the dynamics. Finally, we empirically show that the dictionary outperforms the Max separation function, which had previously been argued to be optimal, and that performance can further be improved by replacing the Euclidean distance kernel by a Manhattan distance kernel. These results are enabled by the generalization capabilities of the dictionary, so a novel criterion is introduced to disentangle memorization from generalization when evaluating associative memory models.

Despite the success of Large Language Models (LLMs) on various tasks following human instructions, controlling model generation to follow strict constraints at inference time poses a persistent challenge. In this paper, we introduce Ctrl-G, a neuro-symbolic framework that enables tractable and adaptable control of LLM generation to follow logical constraints reliably. Ctrl-G combines any production-ready LLM with a Hidden Markov Model (HMM), guiding LLM outputs to adhere to logical constraints represented as deterministic finite automata. We show that Ctrl-G, when a TULU2-7B model is coupled with a 2B-parameter HMM, outperforms GPT4 in text editing: on the task of generating text insertions/continuations following logical constraints, our approach achieves over 30% higher satisfaction rate in human evaluation. When applied to medium-size language models (e.g., GPT2-large), Ctrl-G also beats its counterparts on standard benchmarks by large margins. Additionally, as a proof-of-concept study, we use Ctrl-G to assist LLM reasoning on the GSM benchmark, foreshadowing the application of Ctrl-G, as well as other constrained generation approaches, beyond traditional language generation tasks.

Learning from previously collected data via behavioral cloning or offline reinforcement learning (RL) is a powerful recipe for scaling generalist agents by avoiding the need for expensive online learning. Despite strong generalization in some respects, agents are often remarkably brittle to minor visual variations in control-irrelevant factors such as the background or camera viewpoint. In this paper, we present theDeepMind Control Visual Benchmark (DMC-VB), a dataset collected in the DeepMind Control Suite to evaluate the robustness of offline RL agents for solving continuous control tasks from visual input in the presence of visual distractors. In contrast to prior works, our dataset (a) combines locomotion and navigation tasks of varying difficulties, (b) includes static and dynamic visual variations, (c) considers data generated by policies with different skill levels, (d) systematically returns pairs of state and pixel observation, (e) is an order of magnitude larger, and (f) includes tasks with hidden goals. Accompanying our dataset, we propose three benchmarks to evaluate representation learning methods for pretraining, and carry out experiments on several recently proposed methods. First, we find that pretrained representations do not help policy learning on DMC-VB, and we highlight a large representation gap between policies learned on pixel observations and on states. Second, we demonstrate when expert data is limited, policy learning can benefit from representations pretrained on (a) suboptimal data, and (b) tasks with stochastic hidden goals. Our dataset and benchmark code to train and evaluate agents are available at https://github.com/google-deepmind/dmc_vision_benchmark.

Reinforcement learning (RL) excels in optimizing policies for discrete-time Markov decision processes (MDP). However, various systems are inherently continuous in time, making discrete-time MDPs an inexact modeling choice. In many applications, such as greenhouse control or medical treatments, each interaction (measurement or switching of action) involves manual intervention and thus is inherently costly. Therefore, we generally prefer a time-adaptive approach with fewer interactions with the system. In this work, we formalize an RL framework, **T**ime-**a**daptive **Co**ntrol \& **S**ensing (**TaCoS**), that tackles this challenge by optimizing over policies that besides control predict the duration of its application. Our formulation results in an extended MDP that any standard RL algorithm can solve. We demonstrate that state-of-the-art RL algorithms trained on TaCoS drastically reduce the interaction amount over their discrete-time counterpart while retaining the same or improved performance, and exhibiting robustness over discretization frequency. Finally, we propose OTaCoS, an efficient model-based algorithm for our setting. We show that OTaCoS enjoys sublinear regret for systems with sufficiently smooth dynamics and empirically results in further sample-efficiency gains.