We describe a system for deep reinforcement learning of robotic manipulation skills applied to a large-scale real-world task: sorting recyclables and trash in office buildings. Real-world deployment of deep RL policies requires not only effective training algorithms, but the ability to bootstrap real-world training and enable broad generalization. To this end, our system combines scalable deep RL from real-world data with bootstrapping from training in simulation, and incorporates auxiliary inputs from existing computer vision systems as a way to boost generalization to novel objects, while retaining the benefits of end-to-end training. We analyze the tradeoffs of different design decisions in our system, and present a large-scale empirical validation that includes training on real-world data gathered over the course of 24 months of experimentation, across a fleet of 23 robots in three office buildings, with a total training set of 9527 hours of robotic experience. Our final validation also consists of 4800 evaluation trials across 240 waste station configurations, in order to evaluate in detail the impact of the design decisions in our system, the scaling effects of including more real-world data, and the performance of the method on novel objects. The projects website and videos can be found at \href{http://rl-at-scale.github.io}{rl-at-scale.github.io}.
Fine manipulation tasks, such as threading cable ties or slotting a battery, are notoriously difficult for robots because they require precision, careful coordination of contact forces, and closed-loop visual feedback. Performing these tasks typically requires high-end robots, accurate sensors, or careful calibration, which can be expensive and difficult to set up. Can learning enable low-cost and imprecise hardware to perform these fine manipulation tasks? We present a low-cost system that performs end-to-end imitation learning directly from real demonstrations, collected with a custom teleoperation interface. Imitation learning, however, presents its own challenges, particularly in high-precision domains: errors in the policy can compound over time, and human demonstrations can be non-stationary. To address these challenges, we develop a simple yet novel algorithm, Action Chunking with Transformers (ACT), which learns a generative model over action sequences. ACT allows the robot to learn 6 difficult tasks in the real world, such as opening a translucent condiment cup and slotting a battery with 80-90% success, with only 10 minutes worth of demonstrations. Project website: https://tonyzhaozh.github.io/aloha/
Effective offline RL methods require properly handling out-of-distribution actions. Implicit Q-learning (IQL) addresses this by training a Q-function using only dataset actions through a modified Bellman backup. However, it is unclear which policy actually attains the values represented by this implicitly trained Q-function. In this paper, we reinterpret IQL as an actor-critic method by generalizing the critic objective and connecting it to a behavior-regularized implicit actor. This generalization shows how the induced actor balances reward maximization and divergence from the behavior policy, with the specific loss choice determining the nature of this tradeoff. Notably, this actor can exhibit complex and multimodal characteristics, suggesting issues with the conditional Gaussian actor fit with advantage weighted regression (AWR) used in prior methods. Instead, we propose using samples from a diffusion parameterized behavior policy and weights computed from the critic to then importance sampled our intended policy. We introduce Implicit Diffusion Q-learning (IDQL), combining our general IQL critic with the policy extraction method. IDQL maintains the ease of implementation of IQL while outperforming prior offline RL methods and demonstrating robustness to hyperparameters. Code is available at https://github.com/philippe-eecs/IDQL.
Deep reinforcement learning algorithms that learn policies by trial-and-error must learn from limited amounts of data collected by actively interacting with the environment. While many prior works have shown that proper regularization techniques are crucial for enabling data-efficient RL, a general understanding of the bottlenecks in data-efficient RL has remained unclear. Consequently, it has been difficult to devise a universal technique that works well across all domains. In this paper, we attempt to understand the primary bottleneck in sample-efficient deep RL by examining several potential hypotheses such as non-stationarity, excessive action distribution shift, and overfitting. We perform thorough empirical analysis on state-based DeepMind control suite (DMC) tasks in a controlled and systematic way to show that high temporal-difference (TD) error on the validation set of transitions is the main culprit that severely affects the performance of deep RL algorithms, and prior methods that lead to good performance do in fact, control the validation TD error to be low. This observation gives us a robust principle for making deep RL efficient: we can hill-climb on the validation TD error by utilizing any form of regularization techniques from supervised learning. We show that a simple online model selection method that targets the validation TD error is effective across state-based DMC and Gym tasks.
Legged robots have enormous potential in their range of capabilities, from navigating unstructured terrains to high-speed running. However, designing robust controllers for highly agile dynamic motions remains a substantial challenge for roboticists. Reinforcement learning (RL) offers a promising data-driven approach for automatically training such controllers. However, exploration in these high-dimensional, underactuated systems remains a significant hurdle for enabling legged robots to learn performant, naturalistic, and versatile agility skills. We propose a framework for training complex robotic skills by transferring experience from existing controllers to jumpstart learning new tasks. To leverage controllers we can acquire in practice, we design this framework to be flexible in terms of their source -- that is, the controllers may have been optimized for a different objective under different dynamics, or may require different knowledge of the surroundings -- and thus may be highly suboptimal for the target task. We show that our method enables learning complex agile jumping behaviors, navigating to goal locations while walking on hind legs, and adapting to new environments. We also demonstrate that the agile behaviors learned in this way are graceful and safe enough to deploy in the real world.
We present a system that enables an autonomous small-scale RC car to drive aggressively from visual observations using reinforcement learning (RL). Our system, FastRLAP (faster lap), trains autonomously in the real world, without human interventions, and without requiring any simulation or expert demonstrations. Our system integrates a number of important components to make this possible: we initialize the representations for the RL policy and value function from a large prior dataset of other robots navigating in other environments (at low speed), which provides a navigation-relevant representation. From here, a sample-efficient online RL method uses a single low-speed user-provided demonstration to determine the desired driving course, extracts a set of navigational checkpoints, and autonomously practices driving through these checkpoints, resetting automatically on collision or failure. Perhaps surprisingly, we find that with appropriate initialization and choice of algorithm, our system can learn to drive over a variety of racing courses with less than 20 minutes of online training. The resulting policies exhibit emergent aggressive driving skills, such as timing braking and acceleration around turns and avoiding areas which impede the robot's motion, approaching the performance of a human driver using a similar first-person interface over the course of training.
Passive observational data, such as human videos, is abundant and rich in information, yet remains largely untapped by current RL methods. Perhaps surprisingly, we show that passive data, despite not having reward or action labels, can still be used to learn features that accelerate downstream RL. Our approach learns from passive data by modeling intentions: measuring how the likelihood of future outcomes change when the agent acts to achieve a particular task. We propose a temporal difference learning objective to learn about intentions, resulting in an algorithm similar to conventional RL, but which learns entirely from passive data. When optimizing this objective, our agent simultaneously learns representations of states, of policies, and of possible outcomes in an environment, all from raw observational data. Both theoretically and empirically, this scheme learns features amenable for value prediction for downstream tasks, and our experiments demonstrate the ability to learn from many forms of passive data, including cross-embodiment video data and YouTube videos.
Object rearrangement is a challenge for embodied agents because solving these tasks requires generalizing across a combinatorially large set of configurations of entities and their locations. Worse, the representations of these entities are unknown and must be inferred from sensory percepts. We present a hierarchical abstraction approach to uncover these underlying entities and achieve combinatorial generalization from unstructured visual inputs. By constructing a factorized transition graph over clusters of entity representations inferred from pixels, we show how to learn a correspondence between intervening on states of entities in the agent's model and acting on objects in the environment. We use this correspondence to develop a method for control that generalizes to different numbers and configurations of objects, which outperforms current offline deep RL methods when evaluated on simulated rearrangement tasks.
In the real world, some of the most complex settings for learned agents involve interaction with humans, who often exhibit suboptimal, unpredictable behavior due to sophisticated biases. Agents that interact with people in such settings end up influencing the actions that these people take. Our goal in this work is to enable agents to leverage that influence to improve the human's performance in collaborative tasks, as the task unfolds. Unlike prior work, we do not assume online training with people (which tends to be too expensive and unsafe), nor access to a high fidelity simulator of the environment. Our idea is that by taking a variety of previously observed human-human interaction data and labeling it with the task reward, offline reinforcement learning (RL) can learn to combine components of behavior, and uncover actions that lead to more desirable human actions. First, we show that offline RL can learn strategies to influence and improve human behavior, despite those strategies not appearing in the dataset, by utilizing components of diverse, suboptimal interactions. In addition, we demonstrate that offline RL can learn influence that adapts with humans, thus achieving long-term coordination with them even when their behavior changes. We evaluate our proposed method with real people in the Overcooked collaborative benchmark domain, and demonstrate successful improvement in human performance.
Informational parsimony -- i.e., using the minimal information required for a task, -- provides a useful inductive bias for learning representations that achieve better generalization by being robust to noise and spurious correlations. We propose information gating in the pixel space as a way to learn more parsimonious representations. Information gating works by learning masks that capture only the minimal information required to solve a given task. Intuitively, our models learn to identify which visual cues actually matter for a given task. We gate information using a differentiable parameterization of the signal-to-noise ratio, which can be applied to arbitrary values in a network, e.g.~masking out pixels at the input layer. We apply our approach, which we call InfoGating, to various objectives such as: multi-step forward and inverse dynamics, Q-learning, behavior cloning, and standard self-supervised tasks. Our experiments show that learning to identify and use minimal information can improve generalization in downstream tasks -- e.g., policies based on info-gated images are considerably more robust to distracting/irrelevant visual features.