Recommender systems are the primary interface connecting users to a wide variety of online content, and therefore must overcome a number of challenges across the user population in order to serve them equitably. To this end, in 2019 we released RecSim, a configurable platform for authoring simulation environments to facilitate the study of RL algorithms (the de facto standard ML approach for addressing sequential decision problems) in recommender systems. However, as the technology has progressed, it has become increasingly important to address the gap between simulation and real-world applications, ensuring that models are flexible and easily extendible, enabling probabilistic inference of user dynamics, and addressing computational efficiency.
To address these issues, we recently released RecSim NG, the “Next Generation” of simulators for recommender systems research and development. RecSim NG is a response to a set of use cases that have emerged as important challenges in the application of simulation to real-world problems. It addresses the gap between simulation and real-world applications, ensures the models are flexible and easily extendible, enables probabilistic inference of user dynamics, and addresses computational efficiency.
Overview of RecSim NG RecSim NG is a scalable, modular, differentiable simulator implemented in Edward2 and TensorFlow. It offers a powerful, general probabilistic programming language for agent-behavior specification.
RecSim NG significantly expands the modeling capabilities of RecSim in two ways. First, the story API allows the simulation of scenarios where an arbitrary number of actors (e.g., recommenders, content consumers, content producers, advertisers) interact with one another. This enables the flexible modeling of entire recommender ecosystems, as opposed to the usual isolated user-recommender interaction setting. Second, we introduced a library of behavioral building blocks that, much like Keras layers, implement well-known modeling primitives that can be assembled to build complex models quickly. Following the object-oriented paradigm, RecSim NG uses entity patterns to encapsulate shared parameters that govern various agent behaviors, like user satisfaction, and uses templates to define large populations of agents concisely in a way that abstracts agent “individuality” without duplicating invariant behaviors.
Apart from the typical use of simulators to generate Monte Carlo samples, RecSim NG directly enables various other forms of probabilistic reasoning. While domain knowledge and intuition are key to modeling any recommendation problem, the simulation fidelity needed to bridge the so-called “sim2real” gap can only be achieved by calibrating the simulator’s model to observed data. For data-driven simulation, RecSim NG makes it easy to implement various model-learning algorithms, such as expectation-maximization (EM), generative adversarial training, etc.
Also available within RecSim NG are tools for probabilistic inference and latent-variable model learning, backed by automatic differentiation and tracing. RecSim NG exposes a small set of Edward2 program transformations tailored to simulation-specific tasks. Its log-probability module can evaluate the probabilities of trajectories according to the probabilistic graphical model induced by the simulation. This, together with the automatic differentiation provided by the TensorFlow runtime, enables the implementation of maximum-likelihood estimation and model learning within the simulation itself. RecSim NG can readily use the Markov-chain Monte Carlo (MCMC) machinery provided by TensorFlow Probability to power posterior inference and latent-variable model learning. For example, a simulation model that describes how latent user attributes (e.g., preferences, intents, satisfaction) are translated into observational data (e.g., clicks, ratings, comments) can be “run in reverse,” that is, real observational data generated by a recommender system can be used to identify the most likely configuration of latent user attributes, which in turn can be used to assess the quality of the user experience. This allows for a simulation model to be integrated directly into the full data-science and model-development workflow.
Assessing recommender ecosystem health, i.e., the long-term impact of recommendation strategies on aspects such as overall satisfaction, collective fairness, and safety, requires the simulation of large multi-agent systems in order to plausibly reproduce the interactions between the different participants of the ecosystem. This, along with the computational load of probabilistic inference tasks, requires an efficient simulation runtime. For computational performance, RecSim NG offers a TensorFlow-based runtime for running simulations on accelerated hardware. The simulation takes advantage of all optimizations offered by TensorFlow's AutoGraph compiler, including accelerated linear algebra (XLA) if available. The simulation will automatically exploit all available cores on the host machine as well as specialized hardware (if run accordingly), such as Tensor Processing Units (TPUs). The core RecSim NG architecture is back-end independent, enabling applications to be developed within other computational frameworks (such as JAX or PyTorch).
Ecosystem Modeling as an Application To demonstrate the capabilities of RecSim NG, we present a very simplified model of multi-agent interactions among users and content providers in a stylized recommender ecosystem1. The simulation captures the dynamics of a recommender system that mediates the interaction between users and content providers by recommending slates of those providers’ content items to users over time. We adopt a simplified user model whereby each user is characterized by a static, observable “user interest vector.” This vector determines a user’s affinity with a recommended item, which are then used as inputs to a choice model that determines a user’s item selection from a recommended slate. A user’s utility for any selected item is simply their affinity for the item, perturbed with Gaussian noise.
The aim of the recommender is to maximize cumulative user utility, over all users, over a fixed horizon. However, interesting ecosystem effects make this challenging, and emerge because of content provider behavior. Like users, each provider has an “interest vector'' around which the content items it makes available are centered, reflecting that provider’s general expertise or tendencies. Providers have their own incentives for making content available: their utility is measured by the number of their items selected by any user over the recent past. Moreover, providers with higher utility generate or make available a greater number of items, increasing the “catalog” from which users (and the recommender) can choose.
We compare two different recommender policies in this setting. The first is a standard “myopic'' policy that, for any user, always recommends the items that have the greatest predicted affinity for that user. Under such a policy, the behavior of providers has the potential to give rise to “rich-get-richer'' phenomena: providers that initially attract users produce more items at subsequent periods, which increases the odds of attracting even further future engagement. This gradual concentration of available items around “mainstream” content providers has a negative impact on overall user utility over time. The second recommender policy is aware of these provider dynamics, which it counteracts by promoting under-served providers.2 While a simple heuristic, the provider-aware policy increases overall user utility over extended horizons.
The number of agents in the simulation is large and we templatize both users and content providers with reusable modeling blocks offered by RecSim NG. Determining how to execute the simulation in parallel is non-trivial, so it is critical to utilize TF's AutoGraph and other computational optimizations.
Conclusion Our hope is that RecSim NG will make it easier for both researchers and practitioners to develop, train and evaluate novel algorithms for recommender systems, especially algorithms intended to optimize system behavior over extended horizons, capture complex multi-agent interactions and incentives, or both. We are also investigating the release of increasingly realistic user models that can serve as benchmarks for the research community, as well as methods that can facilitate “sim2real” transfer using RecSim NG.
Further details regarding the RecSim NG framework can be found in the associated white paper, while code and colabs/tutorials are available here. A video about RecSim NG presented at RecSys-2020 is shown below:
Acknowledgements We thank our collaborators and early adopters of RᴇᴄSɪᴍ NG, including the other members of the RecSim NG team: Vihan Jain, Eugene Ie, Chris Colby, Nicolas Mayoraz, Hubert Pham, Dustin Tran, Ivan Vendrov and Craig Boutilier.
1 This model is a much simpler version of that presented in this ICML-20 paper. ↩
2 This simple heuristic policy is used only to demonstrate RecSim NG’s capabilities. More sophisticated algorithms that compute policies that explicitly maximize long-term user utility are discussed in this ICML-20 paper. ↩
As robots become more ubiquitous in day-to-day life, the complexity of their interactions with each other and with the environment grows. In a controlled environment, such as a lab, multiple robots can coordinate their actions and efforts through a centralized planner that facilitates communication between individual agents. And while much research has been done to address reliable sensor-informed goal navigation, in many real-world applications aligning goals across independent robotic agents must be done without a centralized planner, which poses non-trivial challenges.
An example of such a challenging decentralized task is the rendezvous task, in which multiple agents must agree upon a time and place at which they can meet, without explicitly communicating with one another. This goal alignment task plays an important role in real world multiagent and human-robot settings, e.g., performing object handovers or determining goals on the fly. Solving the decentralized rendezvous task in this situation depends not just on the obstacles in the environment, but also the policies and dynamics of each agent. Addressing potential miscoordination and dealing with noisy sensor data depends on the agents’ ability to model the motions of other agents as well as their own, and to adapt to diverging intentions while using limited information.
In “Model-based Reinforcement Learning for Decentralized Multiagent Rendezvous”, presented at CoRL 2020, we propose an holistic approach to address the challenges of the decentralized rendezvous task, which we call hierarchical predictive planning (HPP). This is a decentralized, model-based reinforcement learning (RL) system that enables agents to align their goals on the fly in the real world. We evaluate HPP in a mixture of real-world and simulated environments and compare it to several learning-based planning and centralized baselines. In those evaluations, we show that HPP is able to more effectively predict and align trajectories, avoid miscoordinations, and directly transfer to the real world without additional fine-tuning.
Putting Together Prediction, Planning and Control Akin to a standard navigation pipeline, our learning-based system consists of three modules: prediction, planning, and control. Each agent employs the prediction model to learn agent motion and to predict the future positions of itself (the ego-agent) and others based on its own observations (e.g., from LiDAR and team position information) of other agents’ behaviors and navigation patterns. So, each agent learns two prediction models, one for its own motion and one for the other agent. These motion predictors constitute the prediction module, and are used by each agent’s planning module.
The output of the prediction module — the estimate of where each agent, both the ego-agent and the other agents, is most likely to be given the ego-agent’s own sensor observations — is useful information for the planning module, which evaluates different goal locations and maintains a belief distribution over where the team should converge. The belief distribution is periodically updated using evaluations provided by the prediction model. An agent samples from this belief distribution to update the goal to which it should navigate.
The selected goal is passed to the agent’s control module, which is equipped with a pre-trained, imperfect navigation policy that can navigate to a given location in the obstacle-laden environment. The control policy then determines what action the robot should execute.
This process of observing other agents, updating belief distributions and navigating to an updated goal repeats until agents have successfully rendezvoused. While the hierarchical planning and control setup are not unusual, our work closes the loop between the control and planning for decentralized multiagent systems by use of the sensor-informed prediction module.
Training the Prediction Models HPP trains motion predictors in simulation, assuming that each agent is controlled by a hidden, perhaps suboptimal, control policy capable of avoiding obstacles. The key difficulty lies in training prediction models without access to other agents’ sensor observations and control policies.
The predictors are trained via self-supervision. To collect the training data, we randomly place all the agents and obstacles in an environment, and each agent is given a random goal (unknown to other agents). As the agents move toward their respective goals, each agent records the experience — its sensor observations and the poses of all agents (itself and other agents). Next, from the recorded experience, the agent learns a separate predictor for each agent in the team including itself (target agent). The training dataset consists of ego-agent initial sensor observations, target agent’s pose and goal, labeled with future ego-observations and target agent poses. The goal and labels are inferred from the recorded experience.
As a result, the predictors learn temporal causality of the present and future ego-agent’s observations and target agent’s poses, conditioned on the target agent’s assumed goals — in other words the models predict where each agent will be in the future based on the present. The predictor training is done only with the information available to agents at the runtime, and in environments independent from the deployment environments.
Selecting Goals for Alignment A model-based RL planner for each agent uses the learned predictors in the deployment environments to guide the agents towards the rendezvous point. The planner takes into account what it believes the other agents would do when also completing the rendezvous task.
To perform this reasoning, each agent independently samples a series of potential goals and selects the goal that it believes it would be the most likely to succeed. This process effectively simulates a centralized planner for fictitious agents by using the prediction models to predict trajectories of those agents moving to a fixed goal. Conditioned on a proposed goal, the algorithm predicts the poses of the agents in the future, which are generated from sequential roll outs of the prediction models. Each goal is then evaluated by scoring the anticipated system state using the task reward favoring goals that bring agents closer together. We use the cross-entropy method (CEM) to convert these goal evaluations into belief updates over potential rendezvous points. Finally, the agent’s planner selects a goal for itself from this new belief distribution and passes this goal to the agent’s control module.
Results We compare HPP against several baselines — MADDPG (learning-based), RRT (planning) with CEM, and centralized baselines that use heuristics for selecting the agent’s rendezvous point — in a mixture of real-world and simulated environments.
There are two main takeaways from our results. One is that HPP enables agents to predict and align trajectories, avoiding miscoordinations. For example:
Conclusion This work presents HPP, a model-based RL approach for decentralized multiagent coordination. Agents first learn to predict where they and their teammates are going to be from their own sensors and decide and navigate to a common goal. Our experiments demonstrate the method generalizes to new environments and handles miscoordination while making no assumptions about the dynamics of other agents. This may be of interest to the larger multiagent research community as a real-world example of a decentralized task using noisy sensors and imperfect controllers, to the motion planning community as an example of a learning-based planning system that closes the loop between the planner and controller, and to the RL community as an example of model-based RL as feedback in a hierarchical, self-supervised prediction setting.
Acknowledgements This research was done by Rose E. Wang, J. Chase Kew, Dennis Lee, Tsang-Wei Edward Lee, Tingnan Zhang, Brian Ichter, Jie Tan, Aleksandra Faust with special thanks to Michael Everett, Oscar Ramirez and Igor Mordatch for the insightful discussions.
People are able to retrieve the visual information about 3D environments from a picture quite easily — we can identify objects, determine instance sizes, and reconstruct 3D scene layout, all using the limited signals contained in 2D images. This ability is commonly known as the inverse projection problem, which refers to reconstructing the ambiguous mapping from the retinal images to the sources of retinal stimulation. Real-world computer vision applications, such as autonomous driving, heavily rely on these capabilities to localize and identify 3D objects, which require vision models to infer the spatial location, semantic class, and instance label for each 3D point projected to the 2D images. The ability to reconstruct the 3D world from images can be decomposed into two disjoint computer vision tasks: monocular depth estimation (predicting depth from a single image) and video panoptic segmentation (the unification of instance segmentation and semantic segmentation, in the video domain). However, research has generally considered each task separately. Tackling these tasks jointly with a unified computer vision model could result in easier deployment and greater efficiency by sharing computation among multiple tasks.
Driven by the potential value of a model that predicts depth and video panoptic segmentation at the same time, we present “ViP-DeepLab: Learning Visual Perception with Depth-aware Video Panoptic Segmentation”, accepted to CVPR 2021. In this work, we propose a new task, depth-aware video panoptic segmentation, that aims to simultaneously tackle monocular depth estimation and video panoptic segmentation. For the new task, we present two derived datasets accompanied by a new evaluation metric called depth-aware video panoptic quality (DVPQ). This new metric includes the metrics for depth estimation and video panoptic segmentation, requiring a vision model to simultaneously tackle the two sub-tasks. To this end, we extend Panoptic-DeepLab by adding network branches for depth and video predictions to create ViP-DeepLab, a unified model that jointly performs video panoptic segmentation and monocular depth estimation for each pixel on the image plane, and achieves state-of-the-art performance on several academic datasets for the sub-tasks. This video demonstrates the new task and shows the results of ViP-DeepLab.
Overview While Panoptic-DeepLab is able to output semantic segmentation, center prediction, and center regression for a single frame, it lacks the capability of depth estimation and temporally consistent instance ID prediction for multiple frames. However, ViP-DeepLab accomplishes this by performing additional predictions from two consecutive frames as input. The first additional output is depth estimation for the first frame, for which it assigns an estimated depth to each pixel. In addition, ViP-DeepLab also performs center regression for two consecutive frames for only the object centers that appear in the first frame. This process is called center offset prediction, and allows ViP-DeepLab to group all the pixels in the two frames to the same object that appears in the first frame. New instances emerge if they are not grouped to the previously detected instances. This process continues for every two consecutive frames (with one overlapping frame) in a video sequence, stitching panoptic predictions together to form predictions with temporally consistent instance IDs. That is, it stitches together where objects are and how they move in a video scene with time.
Neural Network Design Building on top of Panoptic-DeepLab, ViP-DeepLab additionally contains two prediction branches: (1) a depth prediction branch, and (2) a next-frame instance branch. Specifically, the depth prediction head is a simple design that predicts depth regression for every pixel, while the next-frame instance branch predicts the center offsets for the pixels in the second frame with respect to the centers in the first frame.
Results We have tested ViP-DeepLab on multiple popular benchmarks, including Cityscapes-VPS, KITTI Depth Prediction, and KITTI Multi-Object Tracking and Segmentation (MOTS).
Specifically, ViP-DeepLab achieves state-of-the-art (SOTA) results, significantly outperforming previous methods by 5.1% video panoptic quality (VPQ) on the Cityscapes-VPS test set.
ViP-DeepLab ranks 1st on the KITTI depth prediction benchmark, improving over previous methods by 0.65 SILog (the smaller the better).
Additionally, ViP-DeepLab was also 1st on KITTI MOTS pedestrians and 3rd on KITTI MOTS cars ranked by the metric sMOTSA, and now is 3rd for both pedestrians and cars ranked by a newer metric HOTA.
Finally, we also present two new datasets for the new task, depth-aware video panoptic segmentation, and test ViP-DeepLab on them. We hope our ViP-DeepLab results on these two new datasets will serve as a strong baseline for the community to compare against. The results are shown below.
Conclusion With a simple architecture, ViP-DeepLab achieves state-of-the-art performance on video panoptic segmentation, monocular depth estimation, and multi-object tracking and segmentation. We hope that along with MaX-DeepLab, which proposes an efficient dual-path transformer module that allows for end-to-end image panoptic segmentation, ViP-DeepLab is useful to the community and furthers research into a more holistic understanding of scenes in the real world.
Acknowledgements We would like to thank the support and valuable discussions with Yukun Zhu, Hartwig Adam, and Alan Yuille (co-authors of ViP-DeepLab), as well as Maxwell Collins, and the Mobile Vision team.
We're continuously working to improve the Pixel — making it more helpful, more capable, and more fun — with regular updates, such as the recent V8.2 update to the Camera app. One such improvement (launched on Pixel 5 and Pixel 4a 5G in October) is a feature that operates “under the hood”, HDR+ with Bracketing. This feature works by merging images taken with different exposure times to improve image quality (especially in shadows), resulting in more natural colors, improved details and texture, and reduced noise.
Why Are HDR Scenes Hard to Capture? The original HDR+ burst photography system is the engine behind high-quality mobile photography, which captures a rapid series of deliberately underexposed images, then combines and renders them in a way that preserves detail across the range of tones. But this system had one limitation: scenes with high dynamic range (HDR) like the one below were noisy in the shadows because all images captured are underexposed.
Capturing HDR scenes is difficult because of the physical constraints of image sensors combined with limited signal in the shadows. We can correctly expose either the shadows or the highlights, but not both at the same time.
Photographers sometimes work around these limitations by taking two different exposures and combining them. This approach, known as exposure bracketing, can deliver the best of both worlds, but it is time-consuming to do by hand. It is also challenging in computational photography because it requires:
To avoid these challenges, the original HDR+ system used a different approach to handle high dynamic range scenes.
The Limits of HDR+ The capture strategy used by HDR+ is based on underexposure, which avoids loss of detail in the highlights. While this strategy comes at the expense of noise in the shadows, HDR+ offsets the increased noise through the use of burst photography.
This approach works well for scenes with moderate dynamic range, but breaks down for HDR scenes. To understand why, we need to take a closer look at how two types of noise get into an image.
Noise in Burst Photography One important type of noise is called shot noise, which depends only on the total amount of light captured — the sum of N frames, each with E seconds of exposure time has the same amount of shot noise as a single frame exposed for N × E seconds. If this were the only type of noise present in captured images, burst photography would be as efficient as taking longer exposures. Unfortunately, a second type of noise, read noise, is introduced by the sensor every time a frame is captured. Read noise doesn’t depend on the amount of light captured but instead depends on the number of frames taken — that is, with each frame taken, an additional fixed amount of read noise is added.
This is why using burst photography to reduce total noise isn’t as efficient as simply taking longer exposures: taking multiple frames can reduce the effect of shot noise, but will also increase read noise. Even though read noise increases with the number of frames, it is still possible to reduce the overall noisiness with burst photography, but it becomes less efficient. If one were to break a long exposure into N shorter exposures, the ratio of signal to noise in the final image would be lower because of the additional read noise. In this case, to get back to the signal-to-noise ratio in the single long exposure, one would need to merge N2 short-exposure frames. In the example below, if a long exposure were divided into 12 short exposures, we'd have to capture 144 (12 × 12) short frames to match the signal-to-noise ratio in the shadows! Capturing and processing this many frames would be much more time consuming — burst capture and processing could take over a minute and result in a poor user experience. Instead, with bracketing one can capture both short and long exposures — combining highlight protection and noise reduction.
Solving with Bracketing While the challenges of bracketing prevented the original HDR+ system from using it, incremental improvements since then, plus a recent concentrated effort, have made it possible in the Camera app. To start, adding bracketing to HDR+ required redesigning the capture strategy. Capturing is complicated by zero shutter lag (ZSL), which underpins the fast capture experience on Pixel. With ZSL, the frames displayed in the viewfinder before the shutter press are the frames we use for HDR+ burst merging. For bracketing, we capture an additional long exposure frame after the shutter press, which is not shown in the viewfinder. Note that holding the camera still for half a second after the shutter press to accommodate the long exposure can help improve image quality, even with a typical amount of handshake.
For Night Sight, the capture strategy isn't constrained by the viewfinder — because all frames are captured after the shutter press while the viewfinder is stopped, this mode easily accommodates capturing longer exposure frames. In this case, we capture three long exposures to further reduce noise.
The Merging Algorithm When merging bracketed shots, we choose one of the short frames as the reference frame to avoid potentially clipped highlights and motion blur. All other frames are aligned to this frame before they are merged. This introduces a challenge — for complex scene motion or occluded regions, it is impossible to find exactly matching regions and a naïve merge algorithm would produce ghosting artifacts in these cases.
To address this, we designed a new spatial merge algorithm, similar to the one used for Super Res Zoom, that decides per pixel whether image content should be merged or not. This deghosting is more complicated for frames with different exposures. Long exposure frames have different noise characteristics, clipped highlights, and different amounts of motion blur, which makes comparisons with the short exposure reference frame more difficult. In addition, ghosting artifacts are more visible in bracketed shots, because noise that would otherwise mask these errors is reduced. Despite those challenges, our algorithm is as robust to these issues as the original HDR+ and Super Res Zoom and doesn’t produce ghosting artifacts. At the same time, it merges images 40% faster than its predecessors. Because it merges RAW images early in the photographic pipeline, we were able to achieve all of those benefits while keeping the rest of processing and the signature HDR+ look unchanged. Furthermore, users who prefer to use computational RAW images can take advantage of those image quality and performance improvements.
Bracketing on Pixel HDR+ with Bracketing is available to users of Pixel 4a (5G) and 5 in the default camera, as well as in Night Sight and Portrait modes. For users of Pixel 4 and 4a, the Google Camera app supports bracketing in Night Sight mode. No user interaction is needed to activate HDR+ with Bracketing — depending on the dynamic range of the scene, and the presence of motion, HDR+ with bracketing chooses the best exposures to maximize image quality (examples).
Acknowledgements HDR+ with Bracketing is the result of a collaboration across several teams at Google. The project would not have been possible without the joint efforts of Sam Hasinoff, Dillon Sharlet, Kiran Murthy, Julian Iseringhausen, Mike Milne, Andy Radin, Nicholas Wilson, Navin Sarma, Gabriel Nava, Emily To, Sushil Nath, Alexander Schiffhauer, Isaac Reynolds, Bill Strathearn, Marius Renn, Alex Hong, Jose Ricardo Lima, Bob Hung, Ying Chen Lou, Joy Hsu, Blade Chiu, David Massoud, Jean Hsu, Ellie Yang, and Marc Levoy.
A long-term, overarching goal of research into reinforcement learning (RL) is to design a single general purpose learning algorithm that can solve a wide array of problems. However, because the RL algorithm taxonomy is quite large, and designing new RL algorithms requires extensive tuning and validation, this goal is a daunting one. A possible solution would be to devise a meta-learning method that could design new RL algorithms that generalize to a wide variety of tasks automatically.
In recent years, AutoML has shown great success in automating the design of machine learning components, such as neural networks architectures and model update rules. One example is Neural Architecture Search (NAS), which has been used to develop better neural network architectures for image classification and efficient architectures for running on phones and hardware accelerators. In addition to NAS, AutoML-Zero shows that it’s even possible to learn the entire algorithm from scratch using basic mathematical operations. One common theme in these approaches is that the neural network architecture or the entire algorithm is represented by a graph, and a separate algorithm is used to optimize the graph for certain objectives.
These earlier approaches were designed for supervised learning, in which the overall algorithm is more straightforward. But in RL, there are more components of the algorithm that could be potential targets for design automation (e.g., neural network architectures for agent networks, strategies for sampling from the replay buffer, overall formulation of the loss function), and it is not always clear what the best model update procedure would be to integrate these components. Prior efforts for the automation RL algorithm discovery have focused primarily on model update rules. These approaches learn the optimizer or RL update procedure itself and commonly represent the update rule with a neural network such as an RNN or CNN, which can be efficiently optimized with gradient-based methods. However, these learned rules are not interpretable or generalizable, because the learned weights are opaque and domain specific.
In our paper “Evolving Reinforcement Learning Algorithms”, accepted at ICLR 2021, we show that it’s possible to learn new, analytically interpretable and generalizable RL algorithms by using a graph representation and applying optimization techniques from the AutoML community. In particular, we represent the loss function, which is used to optimize an agent’s parameters over its experience, as a computational graph, and use Regularized Evolution to evolve a population of the computational graphs over a set of simple training environments. This results in increasingly better RL algorithms, and the discovered algorithms generalize to more complex environments, even those with visual observations like Atari games.
RL Algorithm as a Computational Graph Inspired by ideas from NAS, which searches over the space of graphs representing neural network architectures, we meta-learn RL algorithms by representing the loss function of an RL algorithm as a computational graph. In this case, we use a directed acyclic graph for the loss function, with nodes representing inputs, operators, parameters and outputs. For example, in the computational graph for DQN, input nodes include data from the replay buffer, operator nodes include neural network operators and basic math operators, and the output node represents the loss, which will be minimized with gradient descent.
There are a few benefits of such a representation. This representation is expressive enough to define existing algorithms but also new, undiscovered algorithms. It is also interpretable. This graph representation can be analyzed in the same way as human designed RL algorithms, making it more interpretable than approaches that use black box function approximators for the entire RL update procedure. If researchers can understand why a learned algorithm is better, then they can both modify the internal components of the algorithm to improve it and transfer the beneficial components to other problems. Finally, the representation supports general algorithms that can solve a wide variety of problems.
We implemented this representation using the PyGlove library, which conveniently turns the graph into a search space that can be optimized with regularized evolution.
Evolving RL Algorithms We use an evolutionary based approach to optimize the RL algorithms of interest. First, we initialize a population of training agents with randomized graphs. This population of agents is trained in parallel over a set of training environments. The agents first train on a hurdle environment — an easy environment, such as CartPole, intended to quickly weed out poorly performing programs.
If an agent cannot solve the hurdle environment, the training is stopped early with a score of zero. Otherwise the training proceeds to more difficult environments (e.g., Lunar Lander, simple MiniGrid environments, etc.). The algorithm performance is evaluated and used to update the population, where more promising algorithms are further mutated. To reduce the search space, we use a functional equivalence checker which will skip over newly proposed algorithms if they are functionally the same as previously examined algorithms. This loop continues as new mutated candidate algorithms are trained and evaluated. At the end of training, we select the best algorithm and evaluate its performance over a set of unseen test environments.
The population size in the experiments was around 300 agents, and we observed the evolution of good candidate loss functions after 20-50 thousand mutations, requiring about three days of training. We were able to train on CPUs because the training environments were simple, controlling for the computational and energy cost of training. To further control the cost of training, we seeded the initial population with human-designed RL algorithms such as DQN.
Learned Algorithms We highlight two discovered algorithms that exhibit good generalization performance. The first is DQNReg, which builds on DQN by adding a weighted penalty on the Q-values to the normal squared Bellman error. The second learned loss function, DQNClipped, is more complex, although its dominating term has a simple form — the max of the Q-value and the squared Bellman error (modulo a constant). Both algorithms can be viewed as a way to regularize the Q-values. While DQNReg adds a soft constraint, DQNClipped can be interpreted as a kind of constrained optimization that will minimize the Q-values if they become too large. We show that this learned constraint kicks in during the early stage of training when overestimating the Q-values is a potential issue. Once this constraint is satisfied, the loss will instead minimize the original squared Bellman error.
A closer analysis shows that while baselines like DQN commonly overestimate Q-values, our learned algorithms address this issue in different ways. DQNReg underestimates the Q-values, while DQNClipped has similar behavior to double dqn in that it slowly approaches the ground truth without overestimating it.
It’s worth pointing out that these two algorithms consistently emerge when the evolution is seeded with DQN. Learning from scratch, the method rediscovers the TD algorithm. For completeness, we release a dataset of top 1000 performing algorithms discovered during evolution. Curious readers could further investigate the properties of these learned loss functions.
Learned Algorithms Generalization Performance Normally in RL, generalization refers to a trained policy generalizing across tasks. However, in this work we’re interested in algorithmic generalization performance, which means how well an algorithm works over a set of environments. On a set of classical control environments, the learned algorithms can match baselines on the dense reward tasks (CartPole, Acrobot, LunarLander) and outperform DQN on the sparser reward task, MountainCar.
On a set of sparse reward MiniGrid environments, which test a variety of different tasks, we see that DQNReg greatly outperforms baselines on both the training and test environments, in terms of sample efficiency and final performance. In fact, the effect is even more pronounced on the test environments, which vary in size, configuration, and existence of new obstacles, such as lava.
We visualize the performance of normal DDQN vs. the learned algorithm DQNReg on a few MiniGrid environments. The starting location, wall configuration, and object configuration of these environments are randomized at each reset, which requires the agent to generalize instead of simply memorizing the environment. While DDQN often struggles to learn any meaningful behavior, DQNReg can learn the optimal behavior efficiently.
Even on image-based Atari environments we observe improved performance, even though training was on non-image-based environments. This suggests that meta-training on a set of cheap but diverse training environments with a generalizable algorithm representation could enable radical algorithmic generalization.
Conclusion In this post, we’ve discussed learning new interpretable RL algorithms by representing their loss functions as computational graphs and evolving a population of agents over this representation. The computational graph formulation allows researchers to both build upon human-designed algorithms and study the learned algorithms using the same mathematical toolset as the existing algorithms. We analyzed a few of the learned algorithms and can interpret them as a form of entropy regularization to prevent value overestimation. These learned algorithms can outperform baselines and generalize to unseen environments. The top performing algorithms are available for further analytical study.
We hope that future work will extend to more varied RL settings such as actor critic algorithms or offline RL. Furthermore we hope that this work can lead to machine assisted algorithm development where computational meta-learning can help researchers find new directions to pursue and incorporate learned algorithms into their own work.
Acknowledgements We thank our co-authors Daiyi Peng, Esteban Real, Sergey Levine, Quoc V. Le, Honglak Lee, and Aleksandra Faust. We also thank Luke Metz for helpful early discussions and feedback on the paper, Hanjun Dai for early discussions on related research ideas, Xingyou Song, Krzysztof Choromanski, and Kevin Wu for helping with infrastructure, and Jongwook Choi for helping with environment selection. Finally we thank Tom Small for designing animations for this post.
Panoptic segmentation is a computer vision task that unifies semantic segmentation (assigning a class label to each pixel) and instance segmentation (detecting and segmenting each object instance). A core task for real-world applications, panoptic segmentation predicts a set of non-overlapping masks along with their corresponding class labels (i.e., category of object, like "car", "traffic light", "road", etc.) and is generally accomplished using multiple surrogate sub-tasks that approximate (e.g., by using box detection methods) the goals of panoptic segmentation.
Each surrogate sub-task in this proxy tree introduces extra manually-designed modules, such as anchor design rules, box assignment rules, non-maximum suppression (NMS), thing-stuff merging, etc. Although there are good solutions to individual surrogate sub-tasks and modules, undesired artifacts are introduced when these sub-tasks come together in a pipeline for panoptic segmentation, especially in challenging conditions (e.g., two people with similar bounding boxes will trigger NMS, resulting in a missing mask).
Previous efforts, such as DETR, attempted to solve some of these issues by simplifying the box detection sub-task into an end-to-end operation, which is more computationally efficient and results in fewer undesired artifacts. However, the training process still relies heavily on box detection, which does not align with the mask-based definition of panoptic segmentation. Another line of work completely removes boxes from the pipeline, which has the benefit of removing an entire surrogate sub-task along with its associated modules and artifacts. For example, Axial-DeepLab predicts pixel-wise offsets to predefined instance centers, but the surrogate sub-task it uses encounters challenges with highly deformable objects, which have a large variety of shapes (e.g., a cat), or nearby objects with close centers in the image plane, e.g. the image below of a dog seated in a chair.
In “MaX-DeepLab: End-to-End Panoptic Segmentation with Mask Transformers”, to be presented at CVPR 2021, we propose the first fully end-to-end approach for the panoptic segmentation pipeline, directly predicting class-labeled masks by extending the Transformer architecture to this computer vision task. Dubbed MaX-DeepLab for extending Axial-DeepLab with a Mask Xformer, our method employs a dual-path architecture that introduces a global memory path, allowing for direct communication with any convolution layers. As a result, MaX-DeepLab shows a significant 7.1% panoptic quality (PQ) gain in the box-free regime on the challenging COCO dataset, closing the gap between box-based and box-free methods for the first time. MaX-DeepLab achieves the state-of-the-art 51.3% PQ on COCO test-dev set, without test time augmentation.
End-to-End Panoptic Segmentation Inspired by DETR, our model directly predicts a set of non-overlapping masks and their corresponding semantic labels, with output masks and classes that are optimized with a PQ-style objective. Specifically, inspired by the evaluation metric, PQ, which is defined as the recognition quality (whether or not the predicted class is correct) times the segmentation quality (whether the predicted mask is correct), we define a similarity metric between two class-labeled masks in the exact same way. The model is directly trained by maximizing this similarity between ground truth masks and predicted masks via one-to-one matching. This direct modeling of panoptic segmentation enables end-to-end training and inference, removing the hand-coded priors that are necessary in existing box-based and box-free methods.
Dual-Path Transformer Instead of stacking a traditional transformer on top of a convolutional neural network (CNN), we propose a dual-path framework for combining CNNs with transformers. Specifically, we enable any CNN layer to read and write to global memory by using a dual-path transformer block. This proposed block adopts all four types of attention between the CNN-path and the memory-path, and can be inserted anywhere in a CNN, enabling communication with the global memory at any layer. MaX-DeepLab also employs a stacked-hourglass-style decoder that aggregates multi-scale features into a high resolution output. The output is then multiplied with the global memory feature, to form the mask set prediction. The classes for the masks are predicted with another branch of the mask transformer.
Results We evaluate MaX-DeepLab on one of the most challenging panoptic segmentation datasets, COCO, against both of the state-of-the-art box-free (Axial-DeepLab) and box-based (DetectoRS) methods. MaX-DeepLab, without test time augmentation, achieves the state-of-the-art result of 51.3% PQ on the test-dev set.
This result surpasses Axial-DeepLab by 7.1% PQ in the box-free regime and DetectoRS by 1.7% PQ, bridging the gap between box-based and box-free methods for the first time. For a consistent comparison with DETR, we also evaluated a lightweight version of MaX-DeepLab that matches the number of parameters and computations of DETR. The lightweight MaX-DeepLab outperforms DETR by 3.3% PQ on the val set and 3.0% PQ on the test-dev set. In addition, we performed extensive ablation studies and analyses on our end-to-end formulation, model scaling, dual-path architectures, and loss functions. Also the extra-long training schedule of DETR is not necessary for MaX-DeepLab.
As an example in the figure below, MaX-DeepLab correctly segments a dog sitting on a chair. Axial-DeepLab relies on a surrogate sub-task of regressing object center offsets. It fails because the centers of the dog and the chair are close to each other. DetectoRS classifies object bounding boxes, instead of masks, as a surrogate sub-task. It filters out the chair mask because the chair bounding box has a low confidence.
Another example shows how MaX-DeepLab correctly segments images with challenging conditions.
Conclusion We have shown for the first time that panoptic segmentation can be trained end-to-end. MaX-DeepLab directly predicts masks and classes with a mask transformer, removing the need for many hand-designed priors such as object bounding boxes, thing-stuff merging, etc. Equipped with a PQ-style loss and a dual-path transformer, MaX-DeepLab achieves the state-of-the-art result on the challenging COCO dataset, closing the gap between box-based and box-free methods.
Acknowledgements We are thankful to our co-authors, Yukun Zhu, Hartwig Adam, and Alan Yuille. We also thank Maxwell Collins, Sergey Ioffe, Jiquan Ngiam, Siyuan Qiao, Chen Wei, Jieneng Chen, and the Mobile Vision team for the support and valuable discussions.
For general-purpose robots to be most useful, they would need to be able to perform a range of tasks, such as cleaning, maintenance and delivery. But training even a single task (e.g., grasping) using offline reinforcement learning (RL), a trial and error learning method where the agent uses training previously collected data, can take thousands of robot-hours, in addition to the significant engineering needed to enable autonomous operation of a large-scale robotic system. Thus, the computational costs of building general-purpose everyday robots using current robot learning methods becomes prohibitive as the number of tasks grows.
In other large-scale machine learning domains, such as natural language processing and computer vision, a number of strategies have been applied to amortize the effort of learning over multiple skills. For example, pre-training on large natural language datasets can enable few- or zero-shot learning of multiple tasks, such as question answering and sentiment analysis. However, because robots collect their own data, robotic skill learning presents a unique set of opportunities and challenges. Automating this process is a large engineering endeavour, and effectively reusing past robotic data collected by different robots remains an open problem.
Today we present two new advances for robotic RL at scale, MT-Opt, a new multi-task RL system for automated data collection and multi-task RL training, and Actionable Models, which leverages the acquired data for goal-conditioned RL. MT-Opt introduces a scalable data-collection mechanism that is used to collect over 800,000 episodes of various tasks on real robots and demonstrates a successful application of multi-task RL that yields ~3x average improvement over baseline. Additionally, it enables robots to master new tasks quickly through use of its extensive multi-task dataset (new task fine-tuning in <1 day of data collection). Actionable Models enables learning in the absence of specific tasks and rewards by training an implicit model of the world that is also an actionable robotic policy. This drastically increases the number of tasks the robot can perform (via visual goal specification) and enables more efficient learning of downstream tasks.
Large-Scale Multi-Task Data Collection System The cornerstone for both MT-Opt and Actionable Models is the volume and quality of training data. To collect diverse, multi-task data at scale, users need a way to specify tasks, decide for which tasks to collect the data, and finally, manage and balance the resulting dataset. To that end, we create a scalable and intuitive multi-task success detector using data from all of the chosen tasks. The multi-task success is trained using supervised learning to detect the outcome of a given task and it allows users to quickly define new tasks and their rewards. When this success detector is being applied to collect data, it is periodically updated to accommodate distribution shifts caused by various real-world factors, such as varying lighting conditions, changing background surroundings, and novel states that the robots discover.
Second, we simultaneously collect data for multiple distinct tasks across multiple robots by using solutions to easier tasks to effectively bootstrap learning of more complex tasks. This allows training of a policy for the harder tasks and improves the data collected for them. As such, the amount of per-task data and the number of successful episodes for each task grows over time. To further improve the performance, we focus data collection on underperforming tasks, rather than collecting data uniformly across tasks.
This system collected 9600 robot hours of data (from 57 continuous data collection days on seven robots). However, while this data collection strategy was effective at collecting data for a large number of tasks, the success rate and data volume was imbalanced between tasks.
Learning with MT-Opt We address the data collection imbalance by transferring data across tasks and re-balancing the per-task data. The robots generate episodes that are labelled as success or failure for each task and are then copied and shared across other tasks. The balanced batch of episodes is then sent to our multi-task RL training pipeline to train the MT-Opt policy.
MT-Opt uses Q-learning, a popular RL method that learns a function that estimates the future sum of rewards, called the Q-function. The learned policy then picks the action that maximizes this learned Q-function. For multi-task policy training, we specify the task as an extra input to a large Q-learning network (inspired by our previous work on large-scale single-task learning with QT-Opt) and then train all of the tasks simultaneously with offline RL using the entire multi-task dataset. In this way, MT-Opt is able to train on a wide variety of skills that include picking specific objects, placing them into various fixtures, aligning items on a rack, rearranging and covering objects with towels, etc.
Compared to single-task baselines, MT-Opt performs similarly on the tasks that have the most data and significantly improves performance on underrepresented tasks. So, for a generic lifting task, which has the most supporting data, MT-Opt achieved an 89% success rate (compared to 88% for QT-Opt) and achieved a 50% average success rate across rare tasks, compared to 1% with a single-task QT-Opt baseline and 18% using a naïve, multi-task QT-Opt baseline. Using MT-Opt not only enables zero-shot generalization to new but similar tasks, but also can quickly (in about 1 day of data collection on seven robots) be fine-tuned to new, previously unseen tasks. For example, when applied to an unseen towel-covering task, the system achieved a zero-shot success rate of 92% for towel-picking and 79% for object-covering, which wasn’t present in the original dataset.
Learning with Actionable Models While supplying a rigid definition of tasks facilitates autonomous data collection for MT-Opt, it limits the number of learnable behaviors to a fixed set. To enable learning a wider range of tasks from the same data, we use goal-conditioned learning, i.e., learning to reach given goal configurations of a scene in front of the robot, which we specify with goal images. In contrast to explicit model-based methods that learn predictive models of future world observations, or approaches that employ online data collection, this approach learns goal-conditioned policies via offline model-free RL.
To learn to reach any goal state, we perform hindsight relabeling of all trajectories and sub-sequences in our collected dataset and train a goal-conditioned Q-function in a fully offline manner (in contrast to learning online using a fixed set of success examples as in recursive classification). One challenge in this setting is the distributional shift caused by learning only from “positive” hindsight relabeled examples. This we address by employing a conservative strategy to minimize Q-values of unseen actions using artificial negative actions. Furthermore, to enable reaching temporary-extended goals, we introduce a technique for chaining goals across multiple episodes.
Training with Actionable Models allows the system to learn a large repertoire of visually indicated skills, such as object grasping, container placing and object rearrangement. The model is also able to generalize to novel objects and visual objectives not seen in the training data, which demonstrates its ability to learn general functional knowledge about the world. We also show that downstream reinforcement learning tasks can be learned more efficiently by either fine-tuning a pre-trained goal-conditioned model or through a goal-reaching auxiliary objective during training.
Conclusion The results of both MT-Opt and Actionable Models indicate that it is possible to collect and then learn many distinct tasks from large diverse real-robot datasets within a single model, effectively amortizing the cost of learning across many skills. We see this an important step towards general robot learning systems that can be further scaled up to perform many useful services and serve as a starting point for learning downstream tasks.
This post is based on two papers, "MT-Opt: Continuous Multi-Task Robotic Reinforcement Learning at Scale" and "Actionable Models: Unsupervised Offline Reinforcement Learning of Robotic Skills," with additional information and videos on the project websites for MT-Opt and Actionable Models.
Acknowledgements This research was conducted by Dmitry Kalashnikov, Jake Varley, Karol Hausman, Yevgen Chebotar, Ben Swanson, Rico Jonschkowski, Chelsea Finn, Sergey Levine, Yao Lu, Alex Irpan, Ben Eysenbach, Ryan Julian and Ted Xiao. We’d like to give special thanks to Josh Weaver, Noah Brown, Khem Holden, Linda Luu and Brandon Kinman for their robot operation support; Anthony Brohan for help with distributed learning and testing infrastructure; Tom Small for help with videos and project media; Julian Ibarz, Kanishka Rao, Vikas Sindhwani and Vincent Vanhoucke for their support; Tuna Toksoz and Garrett Peake for improving the bin reset mechanisms; Satoshi Kataoka, Michael Ahn, and Ken Oslund for help with the underlying control stack, and the rest of the Robotics at Google team for their overall support and encouragement. All the above contributions were incredibly enabling for this research.
