Offline Q-Learning on Diverse Multi-Task Data Both Scales And Generalizes

The potential of offline reinforcement learning (RL) is that high-capacity models trained on large, heterogeneous datasets can lead to agents that generalize broadly, analogously to similar advances in vision and NLP. However, recent works argue that offline RL methods encounter unique challenges to scaling up model capacity. Drawing on the learnings from these works, we re-examine previous design choices and find that with appropriate choices: ResNets, cross-entropy based distributional backups, and feature normalization, offline Q-learning algorithms exhibit strong performance that scales with model capacity. Using multi-task Atari as a testbed for scaling and generalization, we train a single policy on 40 games with near-human performance using up-to 80 million parameter networks, finding that model performance scales favorably with capacity. In contrast to prior work, we extrapolate beyond dataset performance even when trained entirely on a large (400M transitions) but highly suboptimal dataset (51% human-level performance). Compared to return-conditioned supervised approaches, offline Q-learning scales similarly with model capacity and has better performance, especially when the dataset is suboptimal. Finally, we show that offline Q-learning with a diverse dataset is sufficient to learn powerful representations that facilitate rapid transfer to novel games and fast online learning on new variations of a training game, improving over existing state-of-the-art representation learning approaches.

Data-Driven Offline Decision-Making via Invariant Representation Learning

The goal in offline data-driven decision-making is synthesize decisions that optimize a black-box utility function, using a previously-collected static dataset, with no active interaction. These problems appear in many forms: offline reinforcement learning (RL), where we must produce actions that optimize the long-term reward, bandits from logged data, where the goal is to determine the correct arm, and offline model-based optimization (MBO) problems, where we must find the optimal design provided access to only a static dataset. A key challenge in all these settings is distributional shift: when we optimize with respect to the input into a model trained from offline data, it is easy to produce an out-of-distribution (OOD) input that appears erroneously good. In contrast to prior approaches that utilize pessimism or conservatism to tackle this problem, in this paper, we formulate offline data-driven decision-making as domain adaptation, where the goal is to make accurate predictions for the value of optimized decisions ("target domain"), when training only on the dataset ("source domain"). This perspective leads to invariant objective models (IOM), our approach for addressing distributional shift by enforcing invariance between the learned representations of the training dataset and optimized decisions. In IOM, if the optimized decisions are too different from the training dataset, the representation will be forced to lose much of the information that distinguishes good designs from bad ones, making all choices seem mediocre. Critically, when the optimizer is aware of this representational tradeoff, it should choose not to stray too far from the training distribution, leading to a natural trade-off between distributional shift and learning performance.

Robotic Skill Acquisition via Instruction Augmentation with Vision-Language Models

In recent years, much progress has been made in learning robotic manipulation policies that follow natural language instructions. Such methods typically learn from corpora of robot-language data that was either collected with specific tasks in mind or expensively re-labelled by humans with rich language descriptions in hindsight. Recently, large-scale pretrained vision-language models (VLMs) like CLIP or ViLD have been applied to robotics for learning representations and scene descriptors. Can these pretrained models serve as automatic labelers for robot data, effectively importing Internet-scale knowledge into existing datasets to make them useful even for tasks that are not reflected in their ground truth annotations? To accomplish this, we introduce Data-driven Instruction Augmentation for Language-conditioned control (DIAL): we utilize semi-supervised language labels leveraging the semantic understanding of CLIP to propagate knowledge onto large datasets of unlabelled demonstration data and then train language-conditioned policies on the augmented datasets. This method enables cheaper acquisition of useful language descriptions compared to expensive human labels, allowing for more efficient label coverage of large-scale datasets. We apply DIAL to a challenging real-world robotic manipulation domain where 96.5% of the 80,000 demonstrations do not contain crowd-sourced language annotations. DIAL enables imitation learning policies to acquire new capabilities and generalize to 60 novel instructions unseen in the original dataset.

Offline RL With Realistic Datasets: Heteroskedasticity and Support Constraints

Offline reinforcement learning (RL) learns policies entirely from static datasets, thereby avoiding the challenges associated with online data collection. Practical applications of offline RL will inevitably require learning from datasets where the variability of demonstrated behaviors changes non-uniformly across the state space. For example, at a red light, nearly all human drivers behave similarly by stopping, but when merging onto a highway, some drivers merge quickly, efficiently, and safely, while many hesitate or merge dangerously. Both theoretically and empirically, we show that typical offline RL methods, which are based on distribution constraints fail to learn from data with such non-uniform variability, due to the requirement to stay close to the behavior policy to the same extent across the state space. Ideally, the learned policy should be free to choose per state how closely to follow the behavior policy to maximize long-term return, as long as the learned policy stays within the support of the behavior policy. To instantiate this principle, we reweight the data distribution in conservative Q-learning (CQL) to obtain an approximate support constraint formulation. The reweighted distribution is a mixture of the current policy and an additional policy trained to mine poor actions that are likely under the behavior policy. Our method, CQL (ReDS), is simple, theoretically motivated, and improves performance across a wide range of offline RL problems in Atari games, navigation, and pixel-based manipulation.

Dual Generator Offline Reinforcement Learning

In offline RL, constraining the learned policy to remain close to the data is essential to prevent the policy from outputting out-of-distribution (OOD) actions with erroneously overestimated values. In principle, generative adversarial networks (GAN) can provide an elegant solution to do so, with the discriminator directly providing a probability that quantifies distributional shift. However, in practice, GAN-based offline RL methods have not performed as well as alternative approaches, perhaps because the generator is trained to both fool the discriminator and maximize return -- two objectives that can be at odds with each other. In this paper, we show that the issue of conflicting objectives can be resolved by training two generators: one that maximizes return, with the other capturing the ``remainder'' of the data distribution in the offline dataset, such that the mixture of the two is close to the behavior policy. We show that not only does having two generators enable an effective GAN-based offline RL method, but also approximates a support constraint, where the policy does not need to match the entire data distribution, but only the slice of the data that leads to high long term performance. We name our method DASCO, for Dual-Generator Adversarial Support Constrained Offline RL. On benchmark tasks that require learning from sub-optimal data, DASCO significantly outperforms prior methods that enforce distribution constraint.

Adversarial Policies Beat Professional-Level Go AIs

We attack the state-of-the-art Go-playing AI system, KataGo, by training an adversarial policy that plays against a frozen KataGo victim. Our attack achieves a >99% win-rate against KataGo without search, and a >50% win-rate when KataGo uses enough search to be near-superhuman. To the best of our knowledge, this is the first successful end-to-end attack against a Go AI playing at the level of a top human professional. Notably, the adversary does not win by learning to play Go better than KataGo -- in fact, the adversary is easily beaten by human amateurs. Instead, the adversary wins by tricking KataGo into ending the game prematurely at a point that is favorable to the adversary. Our results demonstrate that even professional-level AI systems may harbor surprising failure modes. See https://goattack.alignmentfund.org/ for example games.

Learning on the Job: Self-Rewarding Offline-to-Online Finetuning for Industrial Insertion of Novel Connectors from Vision

Learning-based methods in robotics hold the promise of generalization, but what can be done if a learned policy does not generalize to a new situation? In principle, if an agent can at least evaluate its own success (i.e., with a reward classifier that generalizes well even when the policy does not), it could actively practice the task and finetune the policy in this situation. We study this problem in the setting of industrial insertion tasks, such as inserting connectors in sockets and setting screws. Existing algorithms rely on precise localization of the connector or socket and carefully managed physical setups, such as assembly lines, to succeed at the task. But in unstructured environments such as homes or even some industrial settings, robots cannot rely on precise localization and may be tasked with previously unseen connectors. Offline reinforcement learning on a variety of connector insertion tasks is a potential solution, but what if the robot is tasked with inserting previously unseen connector? In such a scenario, we will still need methods that can robustly solve such tasks with online practice. One of the main observations we make in this work is that, with a suitable representation learning and domain generalization approach, it can be significantly easier for the reward function to generalize to a new but structurally similar task (e.g., inserting a new type of connector) than for the policy. This means that a learned reward function can be used to facilitate the finetuning of the robot's policy in situations where the policy fails to generalize in zero shot, but the reward function generalizes successfully. We show that such an approach can be instantiated in the real world, pretrained on 50 different connectors, and successfully finetuned to new connectors via the learned reward function. Videos can be viewed at https://sites.google.com/view/learningonthejob

FCM: Forgetful Causal Masking Makes Causal Language Models Better Zero-Shot Learners

Large language models (LLM) trained using the next-token-prediction objective, such as GPT3 and PaLM, have revolutionized natural language processing in recent years by showing impressive zero-shot and few-shot capabilities across a wide range of tasks. In this work, we propose a simple technique that significantly boosts the performance of LLMs without adding computational cost. Our key observation is that, by performing the next token prediction task with randomly selected past tokens masked out, we can improve the quality of the learned representations for downstream language understanding tasks. We hypothesize that randomly masking past tokens prevents over-attending to recent tokens and encourages attention to tokens in the distant past. By randomly masking input tokens in the PaLM model, we show that we can significantly improve 1B and 8B PaLM's zero-shot performance on the SuperGLUE benchmark from 55.7 to 59.2 and from 61.6 to 64.0, respectively. Our largest 8B model matches the score of PaLM with an average score of 64, despite the fact that PaLM is trained on a much larger dataset (780B tokens) of high-quality conversation and webpage data, while ours is trained on the smaller C4 dataset (180B tokens). Experimental results show that our method also improves PaLM's zero and few-shot performance on a diverse suite of tasks, including commonsense reasoning, natural language inference and cloze completion. Moreover, we show that our technique also helps representation learning, significantly improving PaLM's finetuning results.

Unpacking Reward Shaping: Understanding the Benefits of Reward Engineering on Sample Complexity

Reinforcement learning provides an automated framework for learning behaviors from high-level reward specifications, but in practice the choice of reward function can be crucial for good results -- while in principle the reward only needs to specify what the task is, in reality practitioners often need to design more detailed rewards that provide the agent with some hints about how the task should be completed. The idea of this type of ``reward-shaping'' has been often discussed in the literature, and is often a critical part of practical applications, but there is relatively little formal characterization of how the choice of reward shaping can yield benefits in sample complexity. In this work, we build on the framework of novelty-based exploration to provide a simple scheme for incorporating shaped rewards into RL along with an analysis tool to show that particular choices of reward shaping provably improve sample efficiency. We characterize the class of problems where these gains are expected to be significant and show how this can be connected to practical algorithms in the literature. We confirm that these results hold in practice in an experimental evaluation, providing an insight into the mechanisms through which reward shaping can significantly improve the complexity of reinforcement learning while retaining asymptotic performance.

You Only Live Once: Single-Life Reinforcement Learning

Reinforcement learning algorithms are typically designed to learn a performant policy that can repeatedly and autonomously complete a task, usually starting from scratch. However, in many real-world situations, the goal might not be to learn a policy that can do the task repeatedly, but simply to perform a new task successfully once in a single trial. For example, imagine a disaster relief robot tasked with retrieving an item from a fallen building, where it cannot get direct supervision from humans. It must retrieve this object within one test-time trial, and must do so while tackling unknown obstacles, though it may leverage knowledge it has of the building before the disaster. We formalize this problem setting, which we call single-life reinforcement learning (SLRL), where an agent must complete a task within a single episode without interventions, utilizing its prior experience while contending with some form of novelty. SLRL provides a natural setting to study the challenge of autonomously adapting to unfamiliar situations, and we find that algorithms designed for standard episodic reinforcement learning often struggle to recover from out-of-distribution states in this setting. Motivated by this observation, we propose an algorithm, $Q$-weighted adversarial learning (QWALE), which employs a distribution matching strategy that leverages the agent's prior experience as guidance in novel situations. Our experiments on several single-life continuous control problems indicate that methods based on our distribution matching formulation are 20-60% more successful because they can more quickly recover from novel states.