Building generalist agents that can accomplish many goals in rich open-ended environments is one of the research frontiers for reinforcement learning. A key limiting factor for building generalist agents with RL has been the need for a large number of reward functions for achieving different goals. We investigate the feasibility of using off-the-shelf vision-language models, or VLMs, as sources of rewards for reinforcement learning agents. We show how rewards for visual achievement of a variety of language goals can be derived from the CLIP family of models, and used to train RL agents that can achieve a variety of language goals. We showcase this approach in two distinct visual domains and present a scaling trend showing how larger VLMs lead to more accurate rewards for visual goal achievement, which in turn produces more capable RL agents.

We propose Algorithm Distillation (AD), a method for distilling reinforcement learning (RL) algorithms into neural networks by modeling their training histories with a causal sequence model. Algorithm Distillation treats learning to reinforcement learn as an across-episode sequential prediction problem. A dataset of learning histories is generated by a source RL algorithm, and then a causal transformer is trained by autoregressively predicting actions given their preceding learning histories as context. Unlike sequential policy prediction architectures that distill post-learning or expert sequences, AD is able to improve its policy entirely in-context without updating its network parameters. We demonstrate that AD can reinforcement learn in-context in a variety of environments with sparse rewards, combinatorial task structure, and pixel-based observations, and find that AD learns a more data-efficient RL algorithm than the one that generated the source data.

We introduce Contrastive Intrinsic Control (CIC), an unsupervised reinforcement learning (RL) algorithm that maximizes the mutual information between state-transitions and latent skill vectors. CIC utilizes contrastive learning between state-transitions and skills vectors to learn behaviour embeddings and maximizes the entropy of these embeddings as an intrinsic reward to encourage behavioural diversity. We evaluate our algorithm on the Unsupervised RL Benchmark (URLB) in the asymptotic state-based setting, which consists of a long reward-free pre-training phase followed by a short adaptation phase to downstream tasks with extrinsic rewards. We find that CIC improves over prior exploration algorithms in terms of adaptation efficiency to downstream tasks on state-based URLB.

Pre-training deep neural network models using large unlabelled datasets followed by fine-tuning them on small task-specific datasets has emerged as a dominant paradigm in natural language processing (NLP) and computer vision (CV). Despite the widespread success, such a paradigm has remained atypical in reinforcement learning (RL). In this paper, we investigate how we can leverage large reward-free (i.e. task-agnostic) offline datasets of prior interactions to pre-train agents that can then be fine-tuned using a small reward-annotated dataset. To this end, we present Pre-trained Decision Transformer (PDT), a simple yet powerful algorithm for semi-supervised Offline RL. By masking reward tokens during pre-training, the transformer learns to autoregressivley predict actions based on previous state and action context and effectively extracts behaviors present in the dataset. During fine-tuning, rewards are un-masked and the agent learns the set of skills that should be invoked for the desired behavior as per the reward function. We demonstrate the efficacy of this simple and flexible approach on tasks from the D4RL benchmark with limited reward annotations.

Recent progress in deep learning has relied on access to large and diverse datasets. Such data-driven progress has been less evident in offline reinforcement learning (RL), because offline RL data is usually collected to optimize specific target tasks limiting the data's diversity. In this work, we propose Exploratory data for Offline RL (ExORL), a data-centric approach to offline RL. ExORL first generates data with unsupervised reward-free exploration, then relabels this data with a downstream reward before training a policy with offline RL. We find that exploratory data allows vanilla off-policy RL algorithms, without any offline-specific modifications, to outperform or match state-of-the-art offline RL algorithms on downstream tasks. Our findings suggest that data generation is as important as algorithmic advances for offline RL and hence requires careful consideration from the community. Code and data can be found at https://github.com/denisyarats/exorl .

Pre-training large-scale vision and language models (e.g. CLIP) has shown promising results in representation and transfer learning. We investigate the question of how to efficiently adapt these models to downstream tasks. For image classification, linear probes have been the standard for ease of use and efficiency, while for language, other approaches like prompt tuning have emerged. We analyze several fine-tuning methods across a diverse set of image classification tasks across two spectra investigating the amount and similarity of downstream data to that of pretraining one. We find that just tuning LayerNorm parameters is a surprisingly effective baseline across the board. We further demonstrate a simple yet effective strategy that combines LayerNorm-tuning with general fine-tuning methods to improve their performance and benchmark them on few-shot adaption and distribution shift tasks. Finally, we provide an empirical analysis and recommend general recipes for efficient transfer learning of vision and language models. Website at https://sites.google.com/view/adapt-large-scale-models

We introduce Contrastive Intrinsic Control (CIC), an unsupervised reinforcement learning (RL) algorithm that maximizes the mutual information between state-transitions and latent skill vectors. CIC utilizes contrastive learning between state-transitions and skills vectors to learn behaviour embeddings and maximizes the entropy of these embeddings as an intrinsic reward to encourage behavioural diversity. We evaluate our algorithm on the Unsupervised RL Benchmark (URLB) in the asymptotic state-based setting, which consists of a long reward-free pretraining phase followed by a short adaptation phase to downstream tasks with extrinsic rewards. We find that CIC improves over prior exploration algorithms in terms of adaptation efficiency to downstream tasks on state-based URLB. (1)

Recent advances in unsupervised representation learning significantly improved the sample efficiency of training Reinforcement Learning policies in simulated environments. However, similar gains have not yet been seen for real-robot reinforcement learning. In this work, we focus on enabling data-efficient real-robot learning from pixels. We present Contrastive Pre-training and Data Augmentation for Efficient Robotic Learning (CoDER), a method that utilizes data augmentation and unsupervised learning to achieve sample-efficient training of real-robot arm policies from sparse rewards. While contrastive pre-training, data augmentation, demonstrations, and reinforcement learning are alone insufficient for efficient learning, our main contribution is showing that the combination of these disparate techniques results in a simple yet data-efficient method. We show that, given only 10 demonstrations, a single robotic arm can learn sparse-reward manipulation policies from pixels, such as reaching, picking, moving, pulling a large object, flipping a switch, and opening a drawer in just 30 minutes of mean real-world training time. We include videos and code on the project website: https://sites.google.com/view/efficient-robotic-manipulation/home

We introduce Contrastive Intrinsic Control (CIC), an algorithm for unsupervised skill discovery that maximizes the mutual information between state-transitions and latent skill vectors. CIC utilizes contrastive learning between state-transitions and skills to learn behavior embeddings and maximizes the entropy of these embeddings as an intrinsic reward to encourage behavioral diversity. We evaluate our algorithm on the Unsupervised Reinforcement Learning Benchmark, which consists of a long reward-free pre-training phase followed by a short adaptation phase to downstream tasks with extrinsic rewards. CIC substantially improves over prior methods in terms of adaptation efficiency, outperforming prior unsupervised skill discovery methods by 1.79x and the next leading overall exploration algorithm by 1.18x.

A desirable property of autonomous agents is the ability to both solve long-horizon problems and generalize to unseen tasks. Recent advances in data-driven skill learning have shown that extracting behavioral priors from offline data can enable agents to solve challenging long-horizon tasks with reinforcement learning. However, generalization to tasks unseen during behavioral prior training remains an outstanding challenge. To this end, we present Few-shot Imitation with Skill Transition Models (FIST), an algorithm that extracts skills from offline data and utilizes them to generalize to unseen tasks given a few downstream demonstrations. FIST learns an inverse skill dynamics model, a distance function, and utilizes a semi-parametric approach for imitation. We show that FIST is capable of generalizing to new tasks and substantially outperforms prior baselines in navigation experiments requiring traversing unseen parts of a large maze and 7-DoF robotic arm experiments requiring manipulating previously unseen objects in a kitchen.

Temporal information is essential to learning effective policies with Reinforcement Learning (RL). However, current state-of-the-art RL algorithms either assume that such information is given as part of the state space or, when learning from pixels, use the simple heuristic of frame-stacking to implicitly capture temporal information present in the image observations. This heuristic is in contrast to the current paradigm in video classification architectures, which utilize explicit encodings of temporal information through methods such as optical flow and two-stream architectures to achieve state-of-the-art performance. Inspired by leading video classification architectures, we introduce the Flow of Latents for Reinforcement Learning (Flare), a network architecture for RL that explicitly encodes temporal information through latent vector differences. We show that Flare (i) recovers optimal performance in state-based RL without explicit access to the state velocity, solely with positional state information, (ii) achieves state-of-the-art performance on pixel-based challenging continuous control tasks within the DeepMind control benchmark suite, namely quadruped walk, hopper hop, finger turn hard, pendulum swing, and walker run, and is the most sample efficient model-free pixel-based RL algorithm, outperforming the prior model-free state-of-the-art by 1.9X and 1.5X on the 500k and 1M step benchmarks, respectively, and (iv), when augmented over rainbow DQN, outperforms this state-of-the-art level baseline on 5 of 8 challenging Atari games at 100M time step benchmark.

Offline Reinforcement Learning (RL) aims to extract near-optimal policies from imperfect offline data without additional environment interactions. Extracting policies from diverse offline datasets has the potential to expand the range of applicability of RL by making the training process safer, faster, and more streamlined. We investigate how to improve the performance of offline RL algorithms, its robustness to the quality of offline data, as well as its generalization capabilities. To this end, we introduce Offline Model-based RL with Adaptive Behavioral Priors (MABE). Our algorithm is based on the finding that dynamics models, which support within-domain generalization, and behavioral priors, which support cross-domain generalization, are complementary. When combined together, they substantially improve the performance and generalization of offline RL policies. In the widely studied D4RL offline RL benchmark, we find that MABE achieves higher average performance compared to prior model-free and model-based algorithms. In experiments that require cross-domain generalization, we find that MABE outperforms prior methods. Our website is available at https://sites.google.com/berkeley.edu/mabe .

Deep learning models trained on large data sets have been widely successful in both vision and language domains. As state-of-the-art deep learning architectures have continued to grow in parameter count so have the compute budgets and times required to train them, increasing the need for compute-efficient methods that parallelize training. Two common approaches to parallelize the training of deep networks have been data and model parallelism. While useful, data and model parallelism suffer from diminishing returns in terms of compute efficiency for large batch sizes. In this paper, we investigate how to continue scaling compute efficiently beyond the point of diminishing returns for large batches through local parallelism, a framework which parallelizes training of individual layers in deep networks by replacing global backpropagation with truncated layer-wise backpropagation. Local parallelism enables fully asynchronous layer-wise parallelism with a low memory footprint, and requires little communication overhead compared with model parallelism. We show results in both vision and language domains across a diverse set of architectures, and find that local parallelism is particularly effective in the high-compute regime.

Most reinforcement learning (RL) algorithms rely on hand-crafted extrinsic rewards to learn skills. However, crafting a reward function for each skill is not scalable and results in narrow agents that learn reward-specific skills. To alleviate the reliance on reward engineering it is important to develop RL algorithms capable of efficiently acquiring skills with no rewards extrinsic to the agent. While much progress has been made on reward-free exploration in RL, current methods struggle to explore efficiently. Self-play has long been a promising approach for acquiring skills but most successful applications have been in multiagent zero-sum games with extrinsic reward. In this work, we present SelfPlayer, a data-efficient single-agent self-play exploration algorithm. SelfPlayer samples hard but achievable goals from the agent’s past by maximizing a symmetric KL divergence between the visitation distributions of two copies of the agent, Alice and Bob. We show that SelfPlayer outperforms prior leading selfsupervised exploration algorithms such as GoExplore and Curiosity on the data-efficient Atari benchmark.

In an effort to overcome limitations of reward-driven feature learning in deep reinforcement learning (RL) from images, we propose decoupling representation learning from policy learning. To this end, we introduce a new unsupervised learning (UL) task, called Augmented Temporal Contrast (ATC), which trains a convolutional encoder to associate pairs of observations separated by a short time difference, under image augmentations and using a contrastive loss. In online RL experiments, we show that training the encoder exclusively using ATC matches or outperforms end-to-end RL in most environments. Additionally, we benchmark several leading UL algorithms by pre-training encoders on expert demonstrations and using them, with weights frozen, in RL agents; we find that agents using ATC-trained encoders outperform all others. We also train multi-task encoders on data from multiple environments and show generalization to different downstream RL tasks. Finally, we ablate components of ATC, and introduce a new data augmentation to enable replay of (compressed) latent images from pre-trained encoders when RL requires augmentation. Our experiments span visually diverse RL benchmarks in DeepMind Control, DeepMind Lab, and Atari, and our complete code is available at \url{hidden url}.

Off-policy deep reinforcement learning (RL) has been successful in a range of challenging domains. However, standard off-policy RL algorithms can suffer from low signal and even instability in Q-learning because target values are derived from current Q-estimates, which are often noisy. To mitigate the issue, we propose ensemble-based weighted Bellman backups, which re-weight target Q-values based on uncertainty estimates from a Q-ensemble. We empirically observe that the proposed method stabilizes and improves learning on both continuous and discrete control benchmarks. We also specifically investigate the signal-to-noise aspect by studying environments with noisy rewards, and find that weighted Bellman backups significantly outperform standard Bellman backups. Furthermore, since our weighted Bellman backups rely on maintaining an ensemble, we investigate how weighted Bellman backups interact with UCB Exploration. By enforcing the diversity between agents using Bootstrap, we show that these different ideas are largely orthogonal and can be fruitfully integrated, together further improving the performance of existing off-policy RL algorithms, such as Soft Actor-Critic and Rainbow DQN, for both continuous and discrete control tasks on both low-dimensional and high-dimensional environments.

A promising approach to solving challenging long-horizon tasks has been to extract behavior priors (skills) by fitting generative models to large offline datasets of demonstrations. However, such generative models inherit the biases of the underlying data and result in poor and unusable skills when trained on imperfect demonstration data. To better align skill extraction with human intent we present Skill Preferences (SkiP), an algorithm that learns a model over human preferences and uses it to extract human-aligned skills from offline data. After extracting human-preferred skills, SkiP also utilizes human feedback to solve down-stream tasks with RL. We show that SkiP enables a simulated kitchen robot to solve complex multi-step manipulation tasks and substantially outperforms prior leading RL algorithms with human preferences as well as leading skill extraction algorithms without human preferences.

Off-policy deep reinforcement learning (RL) has been successful in a range of challenging domains. However, standard off-policy RL algorithms can suffer from several issues, such as instability in Q-learning and balancing exploration and exploitation. To mitigate these issues, we present SUNRISE, a simple unified ensemble method, which is compatible with various off-policy RL algorithms. SUNRISE integrates two key ingredients: (a) ensemble-based weighted Bellman backups, which re-weight target Q-values based on uncertainty estimates from a Q-ensemble, and (b) an inference method that selects actions using the highest upper-confidence bounds for efficient exploration. By enforcing the diversity between agents using Bootstrap with random initialization, we show that these different ideas are largely orthogonal and can be fruitfully integrated, together further improving the performance of existing off-policy RL algorithms, such as Soft Actor-Critic and Rainbow DQN, for both continuous and discrete control tasks on both low-dimensional and high-dimensional environments. Our training code is available at https://github.com/pokaxpoka/sunrise.

We introduce a framework that abstracts Reinforcement Learning (RL) as a sequence modeling problem. This allows us to draw upon the simplicity and scalability of the Transformer architecture, and associated advances in language modeling such as GPT-x and BERT. In particular, we present Decision Transformer, an architecture that casts the problem of RL as conditional sequence modeling. Unlike prior approaches to RL that fit value functions or compute policy gradients, Decision Transformer simply outputs the optimal actions by leveraging a causally masked Transformer. By conditioning an autoregressive model on the desired return (reward), past states, and actions, our Decision Transformer model can generate future actions that achieve the desired return. Despite its simplicity, Decision Transformer matches or exceeds the performance of state-of-the-art model-free offline RL baselines on Atari, OpenAI Gym, and Key-to-Door tasks.

Deep Reinforcement Learning (RL) has emerged as a powerful paradigm to solve a range of complex yet specific control tasks. Yet training generalist agents that can quickly adapt to new tasks remains an outstanding challenge. Recent advances in unsupervised RL have shown that pre-training RL agents with self-supervised intrinsic rewards can result in efficient adaptation. However, these algorithms have been hard to compare and develop due to the lack of a unified benchmark. To this end, we introduce the Unsupervised Reinforcement Learning Benchmark (URLB). URLB consists of two phases: reward-free pre-training and downstream task adaptation with extrinsic rewards. Building on the DeepMind Control Suite, we provide twelve continuous control tasks from three domains for evaluation and open-source code for eight leading unsupervised RL methods. We find that the implemented baselines make progress but are not able to solve URLB and propose directions for future research.