ICLR 2024 Papers

Training generally capable agents that thoroughly explore their environment and learn new and diverse skills is a long-term goal of robot learning. Quality Diversity Reinforcement Learning (QD-RL) is an emerging research area that blends the best aspects of both fields – Quality Diversity (QD) provides a principled form of exploration and produces collections of behaviorally diverse agents, while Reinforcement Learning (RL) provides a powerful performance improvement operator enabling generalization across tasks and dynamic environments. Existing QD-RL approaches have been constrained to sample efficient, deterministic off- policy RL algorithms and/or evolution strategies and struggle with highly stochastic environments. In this work, we, for the first time, adapt on-policy RL, specifically Proximal Policy Optimization (PPO), to the Differentiable Quality Diversity (DQD) framework and propose several changes that enable efficient optimization and discovery of novel skills on high-dimensional, stochastic robotics tasks. Our new algorithm, Proximal Policy Gradient Arborescence (PPGA), achieves state-of- the-art results, including a 4x improvement in best reward over baselines on the challenging humanoid domain.

Most Neural Radiance Fields (NeRFs) exhibit limited generalization capabilities,which restrict their applicability in representing multiple scenes using a single model. To address this problem, existing generalizable NeRF methods simply condition the model on image features. These methods still struggle to learn precise global representations over diverse scenes since they lack an effective mechanism for interacting among different points and views. In this work, we unveil that 3D implicit representation learning can be significantly improved by mask-based modeling. Specifically, we propose **m**asked **r**ay and **v**iew **m**odeling for generalizable **NeRF** (**MRVM-NeRF**), which is a self-supervised pretraining target to predict complete scene representations from partially masked features along each ray. With this pretraining target, MRVM-NeRF enables better use of correlations across different rays and views as the geometry priors, which thereby strengthens the capability of capturing intricate details within the scenes and boosts the generalization capability across different scenes. Extensive experiments demonstrate the effectiveness of our proposed MRVM-NeRF on both synthetic and real-world datasets, qualitatively and quantitatively. Besides, we also conduct experiments to show the compatibility of our proposed method with various backbones and its superiority under few-shot cases.

Learning representations of geographical space is vital for any machine learning model that integrates geolocated data, spanning application domains such as remote sensing, ecology, or epidemiology. Recent work embeds coordinates using sine and cosine projections based on Double Fourier Sphere (DFS) features. These embeddings assume a rectangular data domain even on global data, which can lead to artifacts, especially at the poles. At the same time, little attention has been paid to the exact design of the neural network architectures with which these functional embeddings are combined. This work proposes a novel location encoder for globally distributed geographic data that combines spherical harmonic basis functions, natively defined on spherical surfaces, with sinusoidal representation networks (SirenNets) that can be interpreted as learned Double Fourier Sphere embedding. We systematically evaluate positional embeddings and neural network architectures across various benchmarks and synthetic evaluation datasets. In contrast to previous approaches that require the combination of both positional encoding and neural networks to learn meaningful representations, we show that both spherical harmonics and sinusoidal representation networks are competitive on their own but set state-of-the-art performances across tasks when combined. The model code and experiments are available at https://github.com/marccoru/locationencoder.

Learning an accurate value function for a given policy is a critical step in solving reinforcement learning (RL) problems. So far, however, the convergence speed and sample complexity performances of most existing policy evaluation algorithms remain unsatisfactory, particularly with non-linear function approximation. This challenge motivates us to develop a new path-integrated primal-dual stochastic gradient (PILOT) method, that is able to achieve a fast convergence speed for RL policy evaluation with nonlinear function approximation. To further alleviate the periodic full gradient evaluation requirement, we further propose an enhanced method with an adaptive-batch adjustment called PILOT$^+$. The main advantages of our methods include: i) PILOT allows the use of {\em{constant}} step sizes and achieves the $\mathcal{O}(1/K)$ convergence rate to first-order stationary points of non-convex policy evaluation problems; ii) PILOT is a generic {\em{single}}-timescale algorithm that is also applicable for solving a large class of non-convex strongly-concave minimax optimization problems; iii) By adaptively adjusting the batch size via historical stochastic gradient information, PILOT$^+$ is more sample-efficient empirically without loss of theoretical convergence rate. Our extensive numerical experiments verify our theoretical findings and showcase the high efficiency of the proposed PILOT and PILOT$^+$ algorithms compared with the state-of-the-art methods.

Replaying data is a principal mechanism underlying the stability and data efficiency of off-policy reinforcement learning (RL). We present an effective yet simple framework to extend the use of replays across multiple experiments, minimally adapting the RL workflow for sizeable improvements in controller performance and research iteration times. At its core, Replay across Experiments (RaE) involves reusing experience from previous experiments to improve exploration and bootstrap learning while reducing required changes to a minimum in comparison to prior work. We empirically show benefits across a number of RL algorithms and challenging control domains spanning both locomotion and manipulation, including hard exploration tasks from egocentric vision. Through comprehensive ablations, we demonstrate robustness to the quality and amount of data available and various hyperparameter choices. Finally, we discuss how our approach can be applied more broadly across research life cycles and can increase resilience by reloading data across random seeds or hyperparameter variations.

Recent works have introduced LEAPS and HPRL, systems that learn latent spaces of domain-specific languages, which are used to define programmatic policies for partially observable Markov decision processes (POMDPs). These systems induce a latent space while optimizing losses such as the behavior loss, which aim to achieve locality in program behavior, meaning that vectors close in the latent space should correspond to similarly behaving programs. In this paper, we show that the programmatic space, induced by the domain-specific language and requiring no training, presents values for the behavior loss similar to those observed in latent spaces presented in previous work. Moreover, algorithms searching in the programmatic space significantly outperform those in LEAPS and HPRL. To explain our results, we measured the "friendliness" of the two spaces to local search algorithms. We discovered that algorithms are more likely to stop at local maxima when searching in the latent space than when searching in the programmatic space. This implies that the optimization topology of the programmatic space, induced by the reward function in conjunction with the neighborhood function, is more conducive to search than that of the latent space. This result provides an explanation for the superior performance in the programmatic space.

In this work, we present an approach to construct a video-based robot policy capable of reliably executing diverse tasks across different robots and environments from few video demonstrations without using any action annotations. Our method leverages images as a task-agnostic representation, encoding both the state and action information, and text as a general representation for specifying robot goals. By synthesizing videos that "hallucinate" robot executing actions and in combination with dense correspondences between frames, our approach can infer the closed-formed action to execute to an environment without the need of any explicit action labels. This unique capability allows us to train the policy solely based on RGB videos and deploy learned policies to various robotic tasks. We demonstrate the efficacy of our approach in learning policies on table-top manipulation and navigation tasks. Additionally, we contribute an open-source framework for efficient video modeling, enabling the training of high-fidelity policy models with four GPUs within a single day.

Are camera poses necessary for multi-view 3D modeling? Existing approaches predominantly assume access to accurate camera poses. While this assumption might hold for dense views, accurately estimating camera poses for sparse views is often elusive. Our analysis reveals that noisy estimated poses lead to degraded performance for existing sparse-view 3D modeling methods. To address this issue, we present LEAP, a novel pose-free approach, therefore challenging the prevailing notion that camera poses are indispensable. LEAP discards pose-based operations and learns geometric knowledge from data. LEAP is equipped with a neural volume, which is shared across scenes and is parameterized to encode geometry and texture priors. For each incoming scene, we update the neural volume by aggregating 2D image features in a feature-similarity-driven manner. The updated neural volume is decoded into the radiance field, enabling novel view synthesis from any viewpoint. On both object-centric and bounded scene-level datasets, we show that LEAP significantly outperforms prior methods when they employ predicted poses from state-of-the-art pose estimators. Notably, LEAP performs on par with prior approaches that use ground-truth poses while running $400\times$ faster than PixelNeRF. We show LEAP generalizes to novel object categories and scenes, and learns knowledge closely resembles epipolar geometry.

Knowledge distillation (KD) is a technique used to transfer knowledge from a larger ''teacher'' model into a smaller ''student'' model. Recent advancements in meta-learning-based knowledge distillation (MetaKD) emphasize that the fine-tuning of teacher models should be aware of the student's need to achieve better knowledge distillation. However, existing MetaKD methods often lack incentives for the teacher model to improve itself. In this study, we introduce MPDistil, a meta-policy distillation technique, that utilizes novel optimization strategies to foster both *collaboration* and *competition* during the fine-tuning of the teacher model in the meta-learning step. Additionally, we propose a curriculum learning framework for the student model in a competitive setup, in which the student model aims to outperform the teacher model by self-training on various tasks. Exhaustive experiments on SuperGLUE and GLUE benchmarks demonstrate the efficacy of MPDistil compared to $20$ conventional KD and advanced MetaKD baselines, showing significant performance enhancements in the student model -- e.g., a distilled 6-layer BERT model outperforms a 12-layer BERT model on five out of six SuperGLUE tasks. Furthermore, MPDistil, while applied to a large language teacher model (DeBERTa-v2-xxlarge), significantly narrows the performance gap of its smaller student counterpart (DeBERTa-12) by just $4.6$% on SuperGLUE. We further demonstrate how higher rewards and customized training curricula strengthen the student model and enhance generalizability.

We introduce a novel task within the field of human motion generation, termed dance accompaniment, which necessitates the generation of responsive movements from a dance partner, the "follower", synchronized with the lead dancer’s movements and the underlying musical rhythm. Unlike existing solo or group dance generation tasks, a duet dance scenario entails a heightened degree of interaction between the two participants, requiring delicate coordination in both pose and position. To support this task, we first build a large-scale and diverse duet interactive dance dataset, DD100, by recording about 117 minutes of professional dancers’ performances. To address the challenges inherent in this task, we propose a GPT based model, Duolando, which autoregressively predicts the subsequent tokenized motion conditioned on the coordinated information of the music, the leader’s and the follower’s movements. To further enhance the GPT’s capabilities of generating stable results on unseen conditions (music and leader motions), we devise an off-policy reinforcement learning strategy that allows the model to explore viable trajectories from out-of-distribution samplings, guided by human-defined rewards. Based on the collected dataset and proposed method, we establish a benchmark with several carefully designed metrics.

Models trained on different datasets can be merged by a weighted-averaging of their parameters, but why does it work and when can it fail? Here, we connect the inaccuracy of weighted-averaging to mismatches in the gradients and propose a new uncertainty-based scheme to improve the performance by reducing the mismatch. The connection also reveals implicit assumptions in other schemes such as averaging, task arithmetic, and Fisher-weighted averaging. Our new method gives consistent improvements for large language models and vision transformers, both in terms of performance and robustness to hyperparameters.

In this paper, we propose a policy gradient method for confounded partially observable Markov decision processes (POMDPs) with continuous state and observation spaces in the offline setting. We first establish a novel identification result to non-parametrically estimate any history-dependent policy gradient under POMDPs using the offline data. The identification enables us to solve a sequence of conditional moment restrictions and adopt the min-max learning procedure with general function approximation for estimating the policy gradient. We then provide a finite-sample non-asymptotic bound for estimating the gradient uniformly over a pre-specified policy class in terms of the sample size, length of horizon, concentratability coefficient and the measure of ill-posedness in solving the conditional moment restrictions. Lastly, by deploying the proposed gradient estimation in the gradient ascent algorithm, we show the global convergence of the proposed algorithm in finding the history-dependent optimal policy under some technical conditions. To the best of our knowledge, this is the first work studying the policy gradient method for POMDPs under the offline setting.

Large-scale protein language models (PLMs), such as the ESM family, have achieved remarkable performance in various downstream tasks related to protein structure and function by undergoing unsupervised training on residue sequences. They have become essential tools for researchers and practitioners in biology. However, a limitation of vanilla PLMs is their lack of explicit consideration for protein structure information, which suggests the potential for further improvement. Motivated by this, we introduce the concept of a ``structure-aware vocabulary" that integrates residue tokens with structure tokens. The structure tokens are derived by encoding the 3D structure of proteins using Foldseek. We then propose SaProt, a large-scale general-purpose PLM trained on an extensive dataset comprising approximately 40 million protein sequences and structures. Through extensive evaluation, our SaProt model surpasses well-established and renowned baselines across 10 significant downstream tasks, demonstrating its exceptional capacity and broad applicability. We have made the code, pre-trained model, and all relevant materials available at https://github.com/westlake-repl/SaProt.

Opponent modeling aims at learning the opponent's behaviors, goals, or beliefs to reduce the uncertainty of the competitive environment and assist decision-making. Existing work has mostly focused on learning opponent models online, which is impractical and inefficient in practical scenarios. To this end, we formalize an Offline Opponent Modeling (OOM) problem with the objective of utilizing pre-collected offline datasets to learn opponent models that characterize the opponent from the viewpoint of the controlled agent, which aids in adapting to the unknown fixed policies of the opponent. Drawing on the promises of the Transformers for decision-making, we introduce a general approach, Transformer Against Opponent (TAO), for OOM. Essentially, TAO tackles the problem by harnessing the full potential of the supervised pre-trained Transformers' in-context learning capabilities. The foundation of TAO lies in three stages: an innovative offline policy embedding learning stage, an offline opponent-aware response policy training stage, and a deployment stage for opponent adaptation with in-context learning. Theoretical analysis establishes TAO's equivalence to Bayesian posterior sampling in opponent modeling and guarantees TAO's convergence in opponent policy recognition. Extensive experiments and ablation studies on competitive environments with sparse and dense rewards demonstrate the impressive performance of TAO. Our approach manifests remarkable prowess for fast adaptation, especially in the face of unseen opponent policies, confirming its in-context learning potency.

Markov Decision Processes (MDPs) are a formal framework for modeling and solving sequential decision-making problems. In finite time horizons such problems are relevant for instance for optimal stopping or specific supply chain problems, but also in the training of large language models. In contrast to infinite horizon MDPs optimal policies are not stationary, policies must be learned for every single epoch. In practice all parameters are often trained simultaneously, ignoring the inherent structure suggested by dynamic programming. This paper introduces a combination of dynamic programming and policy gradient called dynamical policy gradient, where the parameters are trained backwards in time. For the tabular softmax parametrisation we carry out the convergence analysis for simultaneous and dynamic policy gradient towards global optima, both in the exact and sampled gradient settings without regularisation. It turns out that the use of dynamic policy gradient training much better exploits the structure of finite-time problems which is reflected in improved convergence bounds.