Lisa Lee

Anthony Brohan, Noah Brown, Justice Carbajal et al. · stanford

We study how vision-language models trained on Internet-scale data can be incorporated directly into end-to-end robotic control to boost generalization and enable emergent semantic reasoning. Our goal is to enable a single end-to-end trained model to both learn to map robot observations to actions and enjoy the benefits of large-scale pretraining on language and vision-language data from the web. To this end, we propose to co-fine-tune state-of-the-art vision-language models on both robotic trajectory data and Internet-scale vision-language tasks, such as visual question answering. In contrast to other approaches, we propose a simple, general recipe to achieve this goal: in order to fit both natural language responses and robotic actions into the same format, we express the actions as text tokens and incorporate them directly into the training set of the model in the same way as natural language tokens. We refer to such category of models as vision-language-action models (VLA) and instantiate an example of such a model, which we call RT-2. Our extensive evaluation (6k evaluation trials) shows that our approach leads to performant robotic policies and enables RT-2 to obtain a range of emergent capabilities from Internet-scale training. This includes significantly improved generalization to novel objects, the ability to interpret commands not present in the robot training data (such as placing an object onto a particular number or icon), and the ability to perform rudimentary reasoning in response to user commands (such as picking up the smallest or largest object, or the one closest to another object). We further show that incorporating chain of thought reasoning allows RT-2 to perform multi-stage semantic reasoning, for example figuring out which object to pick up for use as an improvised hammer (a rock), or which type of drink is best suited for someone who is tired (an energy drink).

Tongzheng Ren, Tianjun Zhang, Lisa Lee et al. · berkeley

Representation learning often plays a critical role in reinforcement learning by managing the curse of dimensionality. A representative class of algorithms exploits a spectral decomposition of the stochastic transition dynamics to construct representations that enjoy strong theoretical properties in an idealized setting. However, current spectral methods suffer from limited applicability because they are constructed for state-only aggregation and derived from a policy-dependent transition kernel, without considering the issue of exploration. To address these issues, we propose an alternative spectral method, Spectral Decomposition Representation (SPEDER), that extracts a state-action abstraction from the dynamics without inducing spurious dependence on the data collection policy, while also balancing the exploration-versus-exploitation trade-off during learning. A theoretical analysis establishes the sample efficiency of the proposed algorithm in both the online and offline settings. In addition, an experimental investigation demonstrates superior performance over current state-of-the-art algorithms across several benchmarks.

Xinyang Geng, Hao Liu, Lisa Lee et al.

Building scalable models to learn from diverse, multimodal data remains an open challenge. For vision-language data, the dominant approaches are based on contrastive learning objectives that train a separate encoder for each modality. While effective, contrastive learning approaches introduce sampling bias depending on the data augmentations used, which can degrade performance on downstream tasks. Moreover, these methods are limited to paired image-text data, and cannot leverage widely-available unpaired data. In this paper, we investigate whether a large multimodal model trained purely via masked token prediction, without using modality-specific encoders or contrastive learning, can learn transferable representations for downstream tasks. We propose a simple and scalable network architecture, the Multimodal Masked Autoencoder (M3AE), which learns a unified encoder for both vision and language data via masked token prediction. We provide an empirical study of M3AE trained on a large-scale image-text dataset, and find that M3AE is able to learn generalizable representations that transfer well to downstream tasks. Surprisingly, we find that M3AE benefits from a higher text mask ratio (50-90%), in contrast to BERT whose standard masking ratio is 15%, due to the joint training of two data modalities. We also provide qualitative analysis showing that the learned representation incorporates meaningful information from both image and language. Lastly, we demonstrate the scalability of M3AE with larger model size and training time, and its flexibility to train on both paired image-text data as well as unpaired data.

Annie Xie, Lisa Lee, Ted Xiao et al.

What makes generalization hard for imitation learning in visual robotic manipulation? This question is difficult to approach at face value, but the environment from the perspective of a robot can often be decomposed into enumerable factors of variation, such as the lighting conditions or the placement of the camera. Empirically, generalization to some of these factors have presented a greater obstacle than others, but existing work sheds little light on precisely how much each factor contributes to the generalization gap. Towards an answer to this question, we study imitation learning policies in simulation and on a real robot language-conditioned manipulation task to quantify the difficulty of generalization to different (sets of) factors. We also design a new simulated benchmark of 19 tasks with 11 factors of variation to facilitate more controlled evaluations of generalization. From our study, we determine an ordering of factors based on generalization difficulty, that is consistent across simulation and our real robot setup.

Kuang-Huei Lee, Ofir Nachum, Mengjiao Yang et al.

A longstanding goal of the field of AI is a method for learning a highly capable, generalist agent from diverse experience. In the subfields of vision and language, this was largely achieved by scaling up transformer-based models and training them on large, diverse datasets. Motivated by this progress, we investigate whether the same strategy can be used to produce generalist reinforcement learning agents. Specifically, we show that a single transformer-based model - with a single set of weights - trained purely offline can play a suite of up to 46 Atari games simultaneously at close-to-human performance. When trained and evaluated appropriately, we find that the same trends observed in language and vision hold, including scaling of performance with model size and rapid adaptation to new games via fine-tuning. We compare several approaches in this multi-game setting, such as online and offline RL methods and behavioral cloning, and find that our Multi-Game Decision Transformer models offer the best scalability and performance. We release the pre-trained models and code to encourage further research in this direction.

Hao Liu, Lisa Lee, Kimin Lee et al.

Humans are excellent at understanding language and vision to accomplish a wide range of tasks. In contrast, creating general instruction-following embodied agents remains a difficult challenge. Prior work that uses pure language-only models lack visual grounding, making it difficult to connect language instructions with visual observations. On the other hand, methods that use pre-trained multimodal models typically come with divided language and visual representations, requiring designing specialized network architecture to fuse them together. We propose a simple yet effective model for robots to solve instruction-following tasks in vision-based environments. Our \ours method consists of a multimodal transformer that encodes visual observations and language instructions, and a transformer-based policy that predicts actions based on encoded representations. The multimodal transformer is pre-trained on millions of image-text pairs and natural language text, thereby producing generic cross-modal representations of observations and instructions. The transformer-based policy keeps track of the full history of observations and actions, and predicts actions autoregressively. Despite its simplicity, we show that this unified transformer model outperforms all state-of-the-art pre-trained or trained-from-scratch methods in both single-task and multi-task settings. Our model also shows better model scalability and generalization ability than prior work.

Changyeon Kim, Younggyo Seo, Hao Liu et al.

Developing an agent capable of adapting to unseen environments remains a difficult challenge in imitation learning. This work presents Adaptive Return-conditioned Policy (ARP), an efficient framework designed to enhance the agent's generalization ability using natural language task descriptions and pre-trained multimodal encoders. Our key idea is to calculate a similarity between visual observations and natural language instructions in the pre-trained multimodal embedding space (such as CLIP) and use it as a reward signal. We then train a return-conditioned policy using expert demonstrations labeled with multimodal rewards. Because the multimodal rewards provide adaptive signals at each timestep, our ARP effectively mitigates the goal misgeneralization. This results in superior generalization performances even when faced with unseen text instructions, compared to existing text-conditioned policies. To improve the quality of rewards, we also introduce a fine-tuning method for pre-trained multimodal encoders, further enhancing the performance. Video demonstrations and source code are available on the project website: \url{https://sites.google.com/view/2023arp}.

Hao Liu, Xinyang Geng, Lisa Lee et al.

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. We find that our method, Forgetful Causal Masking (FCM), significantly improves both few-shot and finetuning performance of PaLM. We further consider a simple extension, T-FCM, which introduces bidirectional context to causal language model without altering the sequence order, and further improves finetuning performance.

Gemini Team, Petko Georgiev, Ving Ian Lei et al. · deepmind, mila

In this report, we introduce the Gemini 1.5 family of models, representing the next generation of highly compute-efficient multimodal models capable of recalling and reasoning over fine-grained information from millions of tokens of context, including multiple long documents and hours of video and audio. The family includes two new models: (1) an updated Gemini 1.5 Pro, which exceeds the February version on the great majority of capabilities and benchmarks; (2) Gemini 1.5 Flash, a more lightweight variant designed for efficiency with minimal regression in quality. Gemini 1.5 models achieve near-perfect recall on long-context retrieval tasks across modalities, improve the state-of-the-art in long-document QA, long-video QA and long-context ASR, and match or surpass Gemini 1.0 Ultra's state-of-the-art performance across a broad set of benchmarks. Studying the limits of Gemini 1.5's long-context ability, we find continued improvement in next-token prediction and near-perfect retrieval (>99%) up to at least 10M tokens, a generational leap over existing models such as Claude 3.0 (200k) and GPT-4 Turbo (128k). Finally, we highlight real-world use cases, such as Gemini 1.5 collaborating with professionals on completing their tasks achieving 26 to 75% time savings across 10 different job categories, as well as surprising new capabilities of large language models at the frontier; when given a grammar manual for Kalamang, a language with fewer than 200 speakers worldwide, the model learns to translate English to Kalamang at a similar level to a person who learned from the same content.

Gheorghe Comanici, Eric Bieber, Mike Schaekermann et al. · amazon-science, baidu

In this report, we introduce the Gemini 2.X model family: Gemini 2.5 Pro and Gemini 2.5 Flash, as well as our earlier Gemini 2.0 Flash and Flash-Lite models. Gemini 2.5 Pro is our most capable model yet, achieving SoTA performance on frontier coding and reasoning benchmarks. In addition to its incredible coding and reasoning skills, Gemini 2.5 Pro is a thinking model that excels at multimodal understanding and it is now able to process up to 3 hours of video content. Its unique combination of long context, multimodal and reasoning capabilities can be combined to unlock new agentic workflows. Gemini 2.5 Flash provides excellent reasoning abilities at a fraction of the compute and latency requirements and Gemini 2.0 Flash and Flash-Lite provide high performance at low latency and cost. Taken together, the Gemini 2.X model generation spans the full Pareto frontier of model capability vs cost, allowing users to explore the boundaries of what is possible with complex agentic problem solving.

Gemma Team, Thomas Mesnard, Cassidy Hardin et al. · deepmind

This work introduces Gemma, a family of lightweight, state-of-the art open models built from the research and technology used to create Gemini models. Gemma models demonstrate strong performance across academic benchmarks for language understanding, reasoning, and safety. We release two sizes of models (2 billion and 7 billion parameters), and provide both pretrained and fine-tuned checkpoints. Gemma outperforms similarly sized open models on 11 out of 18 text-based tasks, and we present comprehensive evaluations of safety and responsibility aspects of the models, alongside a detailed description of model development. We believe the responsible release of LLMs is critical for improving the safety of frontier models, and for enabling the next wave of LLM innovations.

Gemini Team, Rohan Anil, Sebastian Borgeaud et al.

This report introduces a new family of multimodal models, Gemini, that exhibit remarkable capabilities across image, audio, video, and text understanding. The Gemini family consists of Ultra, Pro, and Nano sizes, suitable for applications ranging from complex reasoning tasks to on-device memory-constrained use-cases. Evaluation on a broad range of benchmarks shows that our most-capable Gemini Ultra model advances the state of the art in 30 of 32 of these benchmarks - notably being the first model to achieve human-expert performance on the well-studied exam benchmark MMLU, and improving the state of the art in every one of the 20 multimodal benchmarks we examined. We believe that the new capabilities of the Gemini family in cross-modal reasoning and language understanding will enable a wide variety of use cases. We discuss our approach toward post-training and deploying Gemini models responsibly to users through services including Gemini, Gemini Advanced, Google AI Studio, and Cloud Vertex AI.

Ken Caluwaerts, Atil Iscen, J. Chase Kew et al.

Animals have evolved various agile locomotion strategies, such as sprinting, leaping, and jumping. There is a growing interest in developing legged robots that move like their biological counterparts and show various agile skills to navigate complex environments quickly. Despite the interest, the field lacks systematic benchmarks to measure the performance of control policies and hardware in agility. We introduce the Barkour benchmark, an obstacle course to quantify agility for legged robots. Inspired by dog agility competitions, it consists of diverse obstacles and a time based scoring mechanism. This encourages researchers to develop controllers that not only move fast, but do so in a controllable and versatile way. To set strong baselines, we present two methods for tackling the benchmark. In the first approach, we train specialist locomotion skills using on-policy reinforcement learning methods and combine them with a high-level navigation controller. In the second approach, we distill the specialist skills into a Transformer-based generalist locomotion policy, named Locomotion-Transformer, that can handle various terrains and adjust the robot's gait based on the perceived environment and robot states. Using a custom-built quadruped robot, we demonstrate that our method can complete the course at half the speed of a dog. We hope that our work represents a step towards creating controllers that enable robots to reach animal-level agility.

Tianwei Ni, Harshit Sikchi, Yufei Wang et al.

Imitation learning is well-suited for robotic tasks where it is difficult to directly program the behavior or specify a cost for optimal control. In this work, we propose a method for learning the reward function (and the corresponding policy) to match the expert state density. Our main result is the analytic gradient of any f-divergence between the agent and expert state distribution w.r.t. reward parameters. Based on the derived gradient, we present an algorithm, f-IRL, that recovers a stationary reward function from the expert density by gradient descent. We show that f-IRL can learn behaviors from a hand-designed target state density or implicitly through expert observations. Our method outperforms adversarial imitation learning methods in terms of sample efficiency and the required number of expert trajectories on IRL benchmarks. Moreover, we show that the recovered reward function can be used to quickly solve downstream tasks, and empirically demonstrate its utility on hard-to-explore tasks and for behavior transfer across changes in dynamics.

Lisa Lee, Benjamin Eysenbach, Ruslan Salakhutdinov et al.

Reinforcement learning (RL) is a powerful framework for learning to take actions to solve tasks. However, in many settings, an agent must winnow down the inconceivably large space of all possible tasks to the single task that it is currently being asked to solve. Can we instead constrain the space of tasks to those that are semantically meaningful? In this work, we introduce a framework for using weak supervision to automatically disentangle this semantically meaningful subspace of tasks from the enormous space of nonsensical "chaff" tasks. We show that this learned subspace enables efficient exploration and provides a representation that captures distance between states. On a variety of challenging, vision-based continuous control problems, our approach leads to substantial performance gains, particularly as the complexity of the environment grows.

Lisa Lee, Benjamin Eysenbach, Emilio Parisotto et al.

Exploration is critical to a reinforcement learning agent's performance in its given environment. Prior exploration methods are often based on using heuristic auxiliary predictions to guide policy behavior, lacking a mathematically-grounded objective with clear properties. In contrast, we recast exploration as a problem of State Marginal Matching (SMM), where we aim to learn a policy for which the state marginal distribution matches a given target state distribution. The target distribution is a uniform distribution in most cases, but can incorporate prior knowledge if available. In effect, SMM amortizes the cost of learning to explore in a given environment. The SMM objective can be viewed as a two-player, zero-sum game between a state density model and a parametric policy, an idea that we use to build an algorithm for optimizing the SMM objective. Using this formalism, we further demonstrate that prior work approximately maximizes the SMM objective, offering an explanation for the success of these methods. On both simulated and real-world tasks, we demonstrate that agents that directly optimize the SMM objective explore faster and adapt more quickly to new tasks as compared to prior exploration methods.

Devendra Singh Chaplot, Lisa Lee, Ruslan Salakhutdinov et al.

Recent efforts on training visual navigation agents conditioned on language using deep reinforcement learning have been successful in learning policies for different multimodal tasks, such as semantic goal navigation and embodied question answering. In this paper, we propose a multitask model capable of jointly learning these multimodal tasks, and transferring knowledge of words and their grounding in visual objects across the tasks. The proposed model uses a novel Dual-Attention unit to disentangle the knowledge of words in the textual representations and visual concepts in the visual representations, and align them with each other. This disentangled task-invariant alignment of representations facilitates grounding and knowledge transfer across both tasks. We show that the proposed model outperforms a range of baselines on both tasks in simulated 3D environments. We also show that this disentanglement of representations makes our model modular, interpretable, and allows for transfer to instructions containing new words by leveraging object detectors.

Maruan Al-Shedivat, Lisa Lee, Ruslan Salakhutdinov et al.

Building agents that can explore their environments intelligently is a challenging open problem. In this paper, we make a step towards understanding how a hierarchical design of the agent's policy can affect its exploration capabilities. First, we design EscapeRoom environments, where the agent must figure out how to navigate to the exit by accomplishing a number of intermediate tasks (\emph{subgoals}), such as finding keys or opening doors. Our environments are procedurally generated and vary in complexity, which can be controlled by the number of subgoals and relationships between them. Next, we propose to measure the complexity of each environment by constructing dependency graphs between the goals and analytically computing \emph{hitting times} of a random walk in the graph. We empirically evaluate Proximal Policy Optimization (PPO) with sparse and shaped rewards, a variation of policy sketches, and a hierarchical version of PPO (called HiPPO) akin to h-DQN. We show that analytically estimated \emph{hitting time} in goal dependency graphs is an informative metric of the environment complexity. We conjecture that the result should hold for environments other than navigation. Finally, we show that solving environments beyond certain level of complexity requires hierarchical approaches.

Lisa Lee, Emilio Parisotto, Devendra Singh Chaplot et al.

Value Iteration Networks (VINs) are effective differentiable path planning modules that can be used by agents to perform navigation while still maintaining end-to-end differentiability of the entire architecture. Despite their effectiveness, they suffer from several disadvantages including training instability, random seed sensitivity, and other optimization problems. In this work, we reframe VINs as recurrent-convolutional networks which demonstrates that VINs couple recurrent convolutions with an unconventional max-pooling activation. From this perspective, we argue that standard gated recurrent update equations could potentially alleviate the optimization issues plaguing VIN. The resulting architecture, which we call the Gated Path Planning Network, is shown to empirically outperform VIN on a variety of metrics such as learning speed, hyperparameter sensitivity, iteration count, and even generalization. Furthermore, we show that this performance gap is consistent across different maze transition types, maze sizes and even show success on a challenging 3D environment, where the planner is only provided with first-person RGB images.

Yohan Jo, Lisa Lee, Shruti Palaskar

There is a great need for technologies that can predict the mortality of patients in intensive care units with both high accuracy and accountability. We present joint end-to-end neural network architectures that combine long short-term memory (LSTM) and a latent topic model to simultaneously train a classifier for mortality prediction and learn latent topics indicative of mortality from textual clinical notes. For topic interpretability, the topic modeling layer has been carefully designed as a single-layer network with constraints inspired by LDA. Experiments on the MIMIC-III dataset show that our models significantly outperform prior models that are based on LDA topics in mortality prediction. However, we achieve limited success with our method for interpreting topics from the trained models by looking at the neural network weights.

Xiaodan Liang, Lisa Lee, Wei Dai et al.

Future frame prediction in videos is a promising avenue for unsupervised video representation learning. Video frames are naturally generated by the inherent pixel flows from preceding frames based on the appearance and motion dynamics in the video. However, existing methods focus on directly hallucinating pixel values, resulting in blurry predictions. In this paper, we develop a dual motion Generative Adversarial Net (GAN) architecture, which learns to explicitly enforce future-frame predictions to be consistent with the pixel-wise flows in the video through a dual-learning mechanism. The primal future-frame prediction and dual future-flow prediction form a closed loop, generating informative feedback signals to each other for better video prediction. To make both synthesized future frames and flows indistinguishable from reality, a dual adversarial training method is proposed to ensure that the future-flow prediction is able to help infer realistic future-frames, while the future-frame prediction in turn leads to realistic optical flows. Our dual motion GAN also handles natural motion uncertainty in different pixel locations with a new probabilistic motion encoder, which is based on variational autoencoders. Extensive experiments demonstrate that the proposed dual motion GAN significantly outperforms state-of-the-art approaches on synthesizing new video frames and predicting future flows. Our model generalizes well across diverse visual scenes and shows superiority in unsupervised video representation learning.

Xiaodan Liang, Lisa Lee, Eric P. Xing

Despite progress in visual perception tasks such as image classification and detection, computers still struggle to understand the interdependency of objects in the scene as a whole, e.g., relations between objects or their attributes. Existing methods often ignore global context cues capturing the interactions among different object instances, and can only recognize a handful of types by exhaustively training individual detectors for all possible relationships. To capture such global interdependency, we propose a deep Variation-structured Reinforcement Learning (VRL) framework to sequentially discover object relationships and attributes in the whole image. First, a directed semantic action graph is built using language priors to provide a rich and compact representation of semantic correlations between object categories, predicates, and attributes. Next, we use a variation-structured traversal over the action graph to construct a small, adaptive action set for each step based on the current state and historical actions. In particular, an ambiguity-aware object mining scheme is used to resolve semantic ambiguity among object categories that the object detector fails to distinguish. We then make sequential predictions using a deep RL framework, incorporating global context cues and semantic embeddings of previously extracted phrases in the state vector. Our experiments on the Visual Relationship Detection (VRD) dataset and the large-scale Visual Genome dataset validate the superiority of VRL, which can achieve significantly better detection results on datasets involving thousands of relationship and attribute types. We also demonstrate that VRL is able to predict unseen types embedded in our action graph by learning correlations on shared graph nodes.

Lisa Lee

Reinforcement learning (RL) is a general and well-known method that a robot can use to learn an optimal control policy to solve a particular task. We would like to build a versatile robot that can learn multiple tasks, but using RL for each of them would be prohibitively expensive in terms of both time and wear-and-tear on the robot. To remedy this problem, we use the Policy Gradient Efficient Lifelong Learning Algorithm (PG-ELLA), an online multi-task RL algorithm that enables the robot to efficiently learn multiple consecutive tasks by sharing knowledge between these tasks to accelerate learning and improve performance. We implemented and evaluated three RL methods--Q-learning, policy gradient RL, and PG-ELLA--on a ground robot whose task is to find a target object in an environment under different surface conditions. In this paper, we discuss our implementations as well as present an empirical analysis of their learning performance.