Model-Based Reinforcement Learning from Pixels with Structured Latent Variable Models


Imagine a robot trying to learn how to stack blocks and push objects using visual inputs from a camera feed. In order to minimize cost and safety concerns, we want our robot to learn these skills with minimal interaction time, but efficient learning from complex sensory inputs such as images is difficult. This work introduces SOLAR, a new model-based reinforcement learning (RL) method that can learn skills – including manipulation tasks on a real Sawyer robot arm – directly from visual inputs with under an hour of interaction. To our knowledge, SOLAR is the most efficient RL method for solving real world image-based robotics tasks.

Our robot learns to stack a Lego block and push a mug onto a coaster with only inputs from a camera pointed at the robot. Each task takes an hour or less of interaction to learn.


Large-Scale Long-Tailed Recognition in an Open World


Existing Computer Vision Setting v.s. Real-World Scenario

One day, an ecologist came to us. He wanted to use modern computer vision techniques to perform automatic animal identification in his wildlife camera trap image datasets. We were so confident because it sounded just like a basic image classification problem. However, we failed. The dataset he provided was extremely long-tailed and open-ended. As usual, when we did not have enough training data, we asked if it was possible to provide more data for the tail classes and just ignore the open classes that might appear in the testing dataset. Unfortunately, collecting more data was not the option. It could take an extremely long time for these ecologists to take photos of rare and secluded animals in the wild. For some endangered animals, they even had to wait for years for one single shot. At the same time, new animal species kept coming in, and old animal species kept leaving. The total class number was never fixed in such a dynamic system. Moreover, the identification of rare and new animals has more conservational values than abundant animals. If we could only do well on the abundant classes, the method would never be practically usable. We tried all possible methods we could think of (data augmentation, sampling techniques, few-shot learning, imbalanced classification, etc.); but none of the existing methods could handle abundant classes, scarce classes and open classes at the same time (Fig. 1).

Figure 1: There exists a considerable gap between the existing computer vision setting and the real-world scenario.


Robots that Learn to Adapt


Figure 1: Our model-based meta reinforcement learning algorithm enables a legged robot to adapt online in the face of an unexpected system malfunction (note the broken front right leg).

Humans have the ability to seamlessly adapt to changes in their environments: adults can learn to walk on crutches in just a few seconds, people can adapt almost instantaneously to picking up an object that is unexpectedly heavy, and children who can walk on flat ground can quickly adapt their gait to walk uphill without having to relearn how to walk. This adaptation is critical for functioning in the real world.


Robots that Learn to Use Improvised Tools


In many animals, tool-use skills emerge from a combination of observational learning and experimentation. For example, by watching one another, chimpanzees can learn how to use twigs to “fish” for insects. Similarly, capuchin monkeys demonstrate the ability to wield sticks as sweeping tools to pull food closer to themselves. While one might wonder whether these are just illustrations of “monkey see, monkey do,” we believe these tool-use abilities indicate a greater level of intelligence.

Left: A chimpanzee fishing for termites. Right: A gorilla using a stick to gather herbs. (source)

The question our new work explores is: can we enable robots to use tools in the same way — through observation and experimentation?

A requisite for performing complex multi-object manipulation tasks, such as those involved in tool use, is an understanding of physical cause-and-effect relationships. Therefore, the ability to predict how one object might interact with another is crucial. Our prior work has investigated how visual predictive models of cause-and-effect can be learned from unsupervised robot interaction with the world. After learning such a model, the robot can plan to accomplish a diverse set of simple tasks, including cloth folding and object arrangement. However, if we consider the more complex interactions that occur in tool-use tasks, such as how a broom can sweep dirt into a dustpan, undirected experimentation isn’t enough.

Hence, taking inspiration from how animals learn, we designed an algorithm that allows robots to learn tool-use skills through a similar paradigm of imitation and interaction. In particular, we show that, with a mix of demonstration data and unsupervised experience, a robot can use novel objects as tools and even improvise tools in the absence of traditional ones. Further, depending on the demands of the task, our method demonstrates the ability to decide whether to use the provided tools. In this post, we will describe how this works.


CVPR 2019 Challenges on Domain Adaptation in Autonomous Driving


We all dream of a future in which autonomous cars can drive us to every corner of the world. Numerous researchers and companies are working day and night to chase this dream by overcoming scientific and technological barriers. One of the greatest challenges we still face is developing machine learning models that can be trained in a local environment and also perform well in new, unseen situations. For example, self-driving cars may utilize perception models to recognize drivable areas from images. Companies in Silicon Valley can build and perfect such a model using large local datasets from the Bay Area for training. However, if the same model were deployed in a snowy area such as Boston, it would likely perform miserably, because it has never seen snow before. Boston, during winter, and Silicon Valley, during any time of the year, can be labeled as separate domains for perception models, since they present clear differences in climate and challenges in perception. In other cases, domains may be much closer in nature, such as a city street and a nearby highway. The process of transferring knowledge and models between different domains in machine learning is called domain adaptation.

A large number of papers on domain adaptation of perception models have appeared in top publishing venues for machine learning and computer vision. However, most of these works focus on image classification and semantic segmentation. Hardly any attention has been paid to instance-level tasks, such as object detection and tracking, even though localization of nearby objects is arguably more important for autonomous driving. To foster the study of domain adaptation of perception models, Berkeley DeepDrive and Didi Chuxing are co-hosting two competitions in CVPR 2019 Workshop on Autonomous Driving. The challenges will focus on domain adaptation of object detection and tracking based on the BDD100K, from Berkeley DeepDrive, and D2-City, from Didi Chuxing, datasets. The domain of BDD100K covers US scenes, while D2-City was collected on China’s streets. The competitions ask participants to transfer object detectors from BDD100K to D2-City and object trackers from D2-city to BDD100K. More information about the challenges can be found on our website and D2-City.

Following our introduction of the BDD100K dataset, we have been busy working to provide more temporal annotations. Above is an example of object tracking annotation, created by our open-source annotation platform Scalabel. Some of the tracking labels are used in the domain adaptation challenge for object tracking. More data will be released this summer. Of course, we also have object tracking at night.


Announcing the BAIR Open Research Commons


The University of California Berkeley Artificial Intelligence Research (BAIR) Lab is pleased to announce the BAIR Open Research Commons, a new industrial affiliate program launched to accelerate cutting-edge AI research. AI research is advancing rapidly in both university and corporate research settings, with existing collaborations already underway driven by individual researcher-to-researcher collaborations. The BAIR Commons is designed to enhance and streamline such collaborative cutting-edge research by students, faculty, and corporate research scholars.

The Commons agreement has been framed with the goal of promoting open research in AI: all on-campus effort, data, and results in the Commons program will be non-exclusive with open publication and open-source code release expected. Fostering an environment for excellence for graduate student research is the primary motivation of the new program: Berkeley students will lead the design of projects in the Commons, and the program of research must be approved by their home departments before a project commences. Students are expected to benefit from collaboration with leading researchers in industrial research labs, as well as the availability of partner resources useful to investigate certain open questions in state-of-the-art AI research. The University will benefit from membership fees paid by partners to participate in the program. The Commons agreement provides for collaborative joint projects between the partners and Berkeley, with intellectual property shared jointly and equally by the parties.

The agreement also provides for joint research “lablets”, which will be embedded collaborative open research spaces inside BAIR’s 27,000 sq. ft. research facility opening this summer in the Berkeley Way West facility on the Berkeley campus. More than a dozen faculty and 120 students will be assigned space in the new lab, with an equal number of visiting positions allocated for researchers from other BAIR labs and for visiting industrial partners.

Initial alliance participants include Amazon, Facebook, Google, Samsung, and Wave Computing. Funding for over twenty joint projects has been committed in the initial launch of the program, which will support both BAIR facilities and research efforts. Over 30 faculty and 200 graduate students and postdocs at Berkeley are affiliated with BAIR. For more information about BAIR or the Commons program please contact

BAIR will occupy the top floor of Berkeley Way West.


Manipulation By Feel


Guiding our fingers while typing, enabling us to nimbly strike a matchstick, and inserting a key in a keyhole all rely on our sense of touch. It has been shown that the sense of touch is very important for dexterous manipulation in humans. Similarly, for many robotic manipulation tasks, vision alone may not be sufficient – often, it may be difficult to resolve subtle details such as the exact position of an edge, shear forces or surface textures at points of contact, and robotic arms and fingers can block the line of sight between a camera and its quarry. Augmenting robots with this crucial sense, however, remains a challenging task.

Our goal is to provide a framework for learning how to perform tactile servoing, which means precisely relocating an object based on tactile information. To provide our robot with tactile feedback, we utilize a custom-built tactile sensor, based on similar principles as the GelSight sensor developed at MIT. The sensor is composed of a deformable, elastomer-based gel, backlit by three colored LEDs, and provides high-resolution RGB images of contact at the gel surface. Compared to other sensors, this tactile sensor sensor naturally provides geometric information in the form of rich visual information from which attributes such as force can be inferred. Previous work using similar sensors has leveraged the this kind of tactile sensor on tasks such as learning how to grasp, improving success rates when grasping a variety of objects.


Assessing Generalization in Deep Reinforcement Learning



We present a benchmark for studying generalization in deep reinforcement learning (RL). Systematic empirical evaluation shows that vanilla deep RL algorithms generalize better than specialized deep RL algorithms designed specifically for generalization. In other words, simply training on varied environments is so far the most effective strategy for generalization. The code can be found at and the full paper is at


Controlling False Discoveries in Large-Scale Experimentation: Challenges and Solutions


“Scientific research has changed the world. Now it needs to change itself.”

- The Economist, 2013

There has been a growing concern about the validity of scientific findings. A multitude of journals, papers and reports have recognized the ever smaller number of replicable scientific studies. In 2016, one of the giants of scientific publishing, Nature, surveyed about 1,500 researchers across many different disciplines, asking for their stand on the status of reproducibility in their area of research. One of the many takeaways to the worrisome results of this survey is the following: 90% of the respondents agreed that there is a reproducibility crisis, and the overall top answer to boosting reproducibility was “better understanding of statistics”. Indeed, many factors contributing to the explosion of irreproducible research stem from the neglect of the fact that statistics is no longer as static as it was in the first half of the 20th century, when statistical hypothesis testing came into prominence as a theoretically rigorous proposal for making valid discoveries with high confidence.


Learning Preferences by Looking at the World


It would be great if we could all have household robots do our chores for us. Chores are tasks that we want done to make our houses cater more to our preferences; they are a way in which we want our house to be different from the way it currently is. However, most “different” states are not very desirable:

Surely our robot wouldn’t be so dumb as to go around breaking stuff when we ask it to clean our house? Unfortunately, AI systems trained with reinforcement learning only optimize features specified in the reward function and are indifferent to anything we might’ve inadvertently left out. Generally, it is easy to get the reward wrong by forgetting to include preferences for things that should stay the same, since we are so used to having these preferences satisfied, and there are so many of them. Consider the room below, and imagine that we want a robot waiter that serves people at the dining table efficiently. We might implement this using a reward function that provides 1 reward whenever the robot serves a dish, and use discounting so that the robot is incentivized to be efficient. What could go wrong with such a reward function? How would we need to modify the reward function to take this into account? Take a minute to think about it.