Andy Zeng

Danny Driess, Fei Xia, Mehdi S. M. Sajjadi et al. · deepmind

Large language models excel at a wide range of complex tasks. However, enabling general inference in the real world, e.g., for robotics problems, raises the challenge of grounding. We propose embodied language models to directly incorporate real-world continuous sensor modalities into language models and thereby establish the link between words and percepts. Input to our embodied language model are multi-modal sentences that interleave visual, continuous state estimation, and textual input encodings. We train these encodings end-to-end, in conjunction with a pre-trained large language model, for multiple embodied tasks including sequential robotic manipulation planning, visual question answering, and captioning. Our evaluations show that PaLM-E, a single large embodied multimodal model, can address a variety of embodied reasoning tasks, from a variety of observation modalities, on multiple embodiments, and further, exhibits positive transfer: the model benefits from diverse joint training across internet-scale language, vision, and visual-language domains. Our largest model, PaLM-E-562B with 562B parameters, in addition to being trained on robotics tasks, is a visual-language generalist with state-of-the-art performance on OK-VQA, and retains generalist language capabilities with increasing scale.

Suvir Mirchandani, Fei Xia, Pete Florence et al. · stanford

We observe that pre-trained large language models (LLMs) are capable of autoregressively completing complex token sequences -- from arbitrary ones procedurally generated by probabilistic context-free grammars (PCFG), to more rich spatial patterns found in the Abstraction and Reasoning Corpus (ARC), a general AI benchmark, prompted in the style of ASCII art. Surprisingly, pattern completion proficiency can be partially retained even when the sequences are expressed using tokens randomly sampled from the vocabulary. These results suggest that without any additional training, LLMs can serve as general sequence modelers, driven by in-context learning. In this work, we investigate how these zero-shot capabilities may be applied to problems in robotics -- from extrapolating sequences of numbers that represent states over time to complete simple motions, to least-to-most prompting of reward-conditioned trajectories that can discover and represent closed-loop policies (e.g., a stabilizing controller for CartPole). While difficult to deploy today for real systems due to latency, context size limitations, and compute costs, the approach of using LLMs to drive low-level control may provide an exciting glimpse into how the patterns among words could be transferred to actions.

Yilun Du, Mengjiao Yang, Pete Florence et al. · mit

We are interested in enabling visual planning for complex long-horizon tasks in the space of generated videos and language, leveraging recent advances in large generative models pretrained on Internet-scale data. To this end, we present video language planning (VLP), an algorithm that consists of a tree search procedure, where we train (i) vision-language models to serve as both policies and value functions, and (ii) text-to-video models as dynamics models. VLP takes as input a long-horizon task instruction and current image observation, and outputs a long video plan that provides detailed multimodal (video and language) specifications that describe how to complete the final task. VLP scales with increasing computation budget where more computation time results in improved video plans, and is able to synthesize long-horizon video plans across different robotics domains: from multi-object rearrangement, to multi-camera bi-arm dexterous manipulation. Generated video plans can be translated into real robot actions via goal-conditioned policies, conditioned on each intermediate frame of the generated video. Experiments show that VLP substantially improves long-horizon task success rates compared to prior methods on both simulated and real robots (across 3 hardware platforms).

Sanjay Subramanian, Medhini Narasimhan, Kushal Khangaonkar et al. · berkeley

We present a framework that formulates visual question answering as modular code generation. In contrast to prior work on modular approaches to VQA, our approach requires no additional training and relies on pre-trained language models (LMs), visual models pre-trained on image-caption pairs, and fifty VQA examples used for in-context learning. The generated Python programs invoke and compose the outputs of the visual models using arithmetic and conditional logic. Our approach improves accuracy on the COVR dataset by at least 3% and on the GQA dataset by roughly 2% compared to the few-shot baseline that does not employ code generation.

Wenlong Huang, Fei Xia, Dhruv Shah et al. · berkeley

Recent progress in large language models (LLMs) has demonstrated the ability to learn and leverage Internet-scale knowledge through pre-training with autoregressive models. Unfortunately, applying such models to settings with embodied agents, such as robots, is challenging due to their lack of experience with the physical world, inability to parse non-language observations, and ignorance of rewards or safety constraints that robots may require. On the other hand, language-conditioned robotic policies that learn from interaction data can provide the necessary grounding that allows the agent to be correctly situated in the real world, but such policies are limited by the lack of high-level semantic understanding due to the limited breadth of the interaction data available for training them. Thus, if we want to make use of the semantic knowledge in a language model while still situating it in an embodied setting, we must construct an action sequence that is both likely according to the language model and also realizable according to grounded models of the environment. We frame this as a problem similar to probabilistic filtering: decode a sequence that both has high probability under the language model and high probability under a set of grounded model objectives. We demonstrate how such grounded models can be obtained across three simulation and real-world domains, and that the proposed decoding strategy is able to solve complex, long-horizon embodiment tasks in a robotic setting by leveraging the knowledge of both models. The project's website can be found at grounded-decoding.github.io.

Wenlong Huang, Fei Xia, Ted Xiao et al.

Recent works have shown how the reasoning capabilities of Large Language Models (LLMs) can be applied to domains beyond natural language processing, such as planning and interaction for robots. These embodied problems require an agent to understand many semantic aspects of the world: the repertoire of skills available, how these skills influence the world, and how changes to the world map back to the language. LLMs planning in embodied environments need to consider not just what skills to do, but also how and when to do them - answers that change over time in response to the agent's own choices. In this work, we investigate to what extent LLMs used in such embodied contexts can reason over sources of feedback provided through natural language, without any additional training. We propose that by leveraging environment feedback, LLMs are able to form an inner monologue that allows them to more richly process and plan in robotic control scenarios. We investigate a variety of sources of feedback, such as success detection, scene description, and human interaction. We find that closed-loop language feedback significantly improves high-level instruction completion on three domains, including simulated and real table top rearrangement tasks and long-horizon mobile manipulation tasks in a kitchen environment in the real world.

Michael Ahn, Anthony Brohan, Noah Brown et al.

Large language models can encode a wealth of semantic knowledge about the world. Such knowledge could be extremely useful to robots aiming to act upon high-level, temporally extended instructions expressed in natural language. However, a significant weakness of language models is that they lack real-world experience, which makes it difficult to leverage them for decision making within a given embodiment. For example, asking a language model to describe how to clean a spill might result in a reasonable narrative, but it may not be applicable to a particular agent, such as a robot, that needs to perform this task in a particular environment. We propose to provide real-world grounding by means of pretrained skills, which are used to constrain the model to propose natural language actions that are both feasible and contextually appropriate. The robot can act as the language model's "hands and eyes," while the language model supplies high-level semantic knowledge about the task. We show how low-level skills can be combined with large language models so that the language model provides high-level knowledge about the procedures for performing complex and temporally-extended instructions, while value functions associated with these skills provide the grounding necessary to connect this knowledge to a particular physical environment. We evaluate our method on a number of real-world robotic tasks, where we show the need for real-world grounding and that this approach is capable of completing long-horizon, abstract, natural language instructions on a mobile manipulator. The project's website and the video can be found at https://say-can.github.io/.

Chenguang Huang, Oier Mees, Andy Zeng et al.

Grounding language to the visual observations of a navigating agent can be performed using off-the-shelf visual-language models pretrained on Internet-scale data (e.g., image captions). While this is useful for matching images to natural language descriptions of object goals, it remains disjoint from the process of mapping the environment, so that it lacks the spatial precision of classic geometric maps. To address this problem, we propose VLMaps, a spatial map representation that directly fuses pretrained visual-language features with a 3D reconstruction of the physical world. VLMaps can be autonomously built from video feed on robots using standard exploration approaches and enables natural language indexing of the map without additional labeled data. Specifically, when combined with large language models (LLMs), VLMaps can be used to (i) translate natural language commands into a sequence of open-vocabulary navigation goals (which, beyond prior work, can be spatial by construction, e.g., "in between the sofa and TV" or "three meters to the right of the chair") directly localized in the map, and (ii) can be shared among multiple robots with different embodiments to generate new obstacle maps on-the-fly (by using a list of obstacle categories). Extensive experiments carried out in simulated and real world environments show that VLMaps enable navigation according to more complex language instructions than existing methods. Videos are available at https://vlmaps.github.io.

Kevin Zakka, Philipp Wu, Laura Smith et al.

Replicating human-like dexterity in robot hands represents one of the largest open problems in robotics. Reinforcement learning is a promising approach that has achieved impressive progress in the last few years; however, the class of problems it has typically addressed corresponds to a rather narrow definition of dexterity as compared to human capabilities. To address this gap, we investigate piano-playing, a skill that challenges even the human limits of dexterity, as a means to test high-dimensional control, and which requires high spatial and temporal precision, and complex finger coordination and planning. We introduce RoboPianist, a system that enables simulated anthropomorphic hands to learn an extensive repertoire of 150 piano pieces where traditional model-based optimization struggles. We additionally introduce an open-sourced environment, benchmark of tasks, interpretable evaluation metrics, and open challenges for future study. Our website featuring videos, code, and datasets is available at https://kzakka.com/robopianist/

Wenhao Yu, Nimrod Gileadi, Chuyuan Fu et al.

Large language models (LLMs) have demonstrated exciting progress in acquiring diverse new capabilities through in-context learning, ranging from logical reasoning to code-writing. Robotics researchers have also explored using LLMs to advance the capabilities of robotic control. However, since low-level robot actions are hardware-dependent and underrepresented in LLM training corpora, existing efforts in applying LLMs to robotics have largely treated LLMs as semantic planners or relied on human-engineered control primitives to interface with the robot. On the other hand, reward functions are shown to be flexible representations that can be optimized for control policies to achieve diverse tasks, while their semantic richness makes them suitable to be specified by LLMs. In this work, we introduce a new paradigm that harnesses this realization by utilizing LLMs to define reward parameters that can be optimized and accomplish variety of robotic tasks. Using reward as the intermediate interface generated by LLMs, we can effectively bridge the gap between high-level language instructions or corrections to low-level robot actions. Meanwhile, combining this with a real-time optimizer, MuJoCo MPC, empowers an interactive behavior creation experience where users can immediately observe the results and provide feedback to the system. To systematically evaluate the performance of our proposed method, we designed a total of 17 tasks for a simulated quadruped robot and a dexterous manipulator robot. We demonstrate that our proposed method reliably tackles 90% of the designed tasks, while a baseline using primitive skills as the interface with Code-as-policies achieves 50% of the tasks. We further validated our method on a real robot arm where complex manipulation skills such as non-prehensile pushing emerge through our interactive system.

Andy Zeng, Maria Attarian, Brian Ichter et al.

Large pretrained (e.g., "foundation") models exhibit distinct capabilities depending on the domain of data they are trained on. While these domains are generic, they may only barely overlap. For example, visual-language models (VLMs) are trained on Internet-scale image captions, but large language models (LMs) are further trained on Internet-scale text with no images (e.g., spreadsheets, SAT questions, code). As a result, these models store different forms of commonsense knowledge across different domains. In this work, we show that this diversity is symbiotic, and can be leveraged through Socratic Models (SMs): a modular framework in which multiple pretrained models may be composed zero-shot i.e., via multimodal-informed prompting, to exchange information with each other and capture new multimodal capabilities, without requiring finetuning. With minimal engineering, SMs are not only competitive with state-of-the-art zero-shot image captioning and video-to-text retrieval, but also enable new applications such as (i) answering free-form questions about egocentric video, (ii) engaging in multimodal assistive dialogue with people (e.g., for cooking recipes) by interfacing with external APIs and databases (e.g., web search), and (iii) robot perception and planning.

Allen Z. Ren, Anushri Dixit, Alexandra Bodrova et al.

Large language models (LLMs) exhibit a wide range of promising capabilities -- from step-by-step planning to commonsense reasoning -- that may provide utility for robots, but remain prone to confidently hallucinated predictions. In this work, we present KnowNo, which is a framework for measuring and aligning the uncertainty of LLM-based planners such that they know when they don't know and ask for help when needed. KnowNo builds on the theory of conformal prediction to provide statistical guarantees on task completion while minimizing human help in complex multi-step planning settings. Experiments across a variety of simulated and real robot setups that involve tasks with different modes of ambiguity (e.g., from spatial to numeric uncertainties, from human preferences to Winograd schemas) show that KnowNo performs favorably over modern baselines (which may involve ensembles or extensive prompt tuning) in terms of improving efficiency and autonomy, while providing formal assurances. KnowNo can be used with LLMs out of the box without model-finetuning, and suggests a promising lightweight approach to modeling uncertainty that can complement and scale with the growing capabilities of foundation models. Website: https://robot-help.github.io

Chenguang Huang, Oier Mees, Andy Zeng et al.

While interacting in the world is a multi-sensory experience, many robots continue to predominantly rely on visual perception to map and navigate in their environments. In this work, we propose Audio-Visual-Language Maps (AVLMaps), a unified 3D spatial map representation for storing cross-modal information from audio, visual, and language cues. AVLMaps integrate the open-vocabulary capabilities of multimodal foundation models pre-trained on Internet-scale data by fusing their features into a centralized 3D voxel grid. In the context of navigation, we show that AVLMaps enable robot systems to index goals in the map based on multimodal queries, e.g., textual descriptions, images, or audio snippets of landmarks. In particular, the addition of audio information enables robots to more reliably disambiguate goal locations. Extensive experiments in simulation show that AVLMaps enable zero-shot multimodal goal navigation from multimodal prompts and provide 50% better recall in ambiguous scenarios. These capabilities extend to mobile robots in the real world - navigating to landmarks referring to visual, audio, and spatial concepts. Videos and code are available at: https://avlmaps.github.io.

Lihan Zha, Yuchen Cui, Li-Heng Lin et al.

Today's robot policies exhibit subpar performance when faced with the challenge of generalizing to novel environments. Human corrective feedback is a crucial form of guidance to enable such generalization. However, adapting to and learning from online human corrections is a non-trivial endeavor: not only do robots need to remember human feedback over time to retrieve the right information in new settings and reduce the intervention rate, but also they would need to be able to respond to feedback that can be arbitrary corrections about high-level human preferences to low-level adjustments to skill parameters. In this work, we present Distillation and Retrieval of Online Corrections (DROC), a large language model (LLM)-based system that can respond to arbitrary forms of language feedback, distill generalizable knowledge from corrections, and retrieve relevant past experiences based on textual and visual similarity for improving performance in novel settings. DROC is able to respond to a sequence of online language corrections that address failures in both high-level task plans and low-level skill primitives. We demonstrate that DROC effectively distills the relevant information from the sequence of online corrections in a knowledge base and retrieves that knowledge in settings with new task or object instances. DROC outperforms other techniques that directly generate robot code via LLMs by using only half of the total number of corrections needed in the first round and requires little to no corrections after two iterations. We show further results, videos, prompts and code on https://sites.google.com/stanford.edu/droc .

Ryan Hoque, Kaushik Shivakumar, Shrey Aeron et al.

Autonomous fabric manipulation is a longstanding challenge in robotics, but evaluating progress is difficult due to the cost and diversity of robot hardware. Using Reach, a cloud robotics platform that enables low-latency remote execution of control policies on physical robots, we present the first systematic benchmarking of fabric manipulation algorithms on physical hardware. We develop 4 novel learning-based algorithms that model expert actions, keypoints, reward functions, and dynamic motions, and we compare these against 4 learning-free and inverse dynamics algorithms on the task of folding a crumpled T-shirt with a single robot arm. The entire lifecycle of data collection, model training, and policy evaluation is performed remotely without physical access to the robot workcell. Results suggest a new algorithm combining imitation learning with analytic methods achieves 84% of human-level performance on the folding task. See https://sites.google.com/berkeley.edu/cloudfolding for all data, code, models, and supplemental material.

Jimmy Wu, Xingyuan Sun, Andy Zeng et al.

We investigate pneumatic non-prehensile manipulation (i.e., blowing) as a means of efficiently moving scattered objects into a target receptacle. Due to the chaotic nature of aerodynamic forces, a blowing controller must (i) continually adapt to unexpected changes from its actions, (ii) maintain fine-grained control, since the slightest misstep can result in large unintended consequences (e.g., scatter objects already in a pile), and (iii) infer long-range plans (e.g., move the robot to strategic blowing locations). We tackle these challenges in the context of deep reinforcement learning, introducing a multi-frequency version of the spatial action maps framework. This allows for efficient learning of vision-based policies that effectively combine high-level planning and low-level closed-loop control for dynamic mobile manipulation. Experiments show that our system learns efficient behaviors for the task, demonstrating in particular that blowing achieves better downstream performance than pushing, and that our policies improve performance over baselines. Moreover, we show that our system naturally encourages emergent specialization between the different subpolicies spanning low-level fine-grained control and high-level planning. On a real mobile robot equipped with a miniature air blower, we show that our simulation-trained policies transfer well to a real environment and can generalize to novel objects.

Yulong Li, Andy Zeng, Shuran Song

Most successes in autonomous robotic assembly have been restricted to single target or category. We propose to investigate general part assembly, the task of creating novel target assemblies with unseen part shapes. As a fundamental step to a general part assembly system, we tackle the task of determining the precise poses of the parts in the target assembly, which we we term ``rearrangement planning''. We present General Part Assembly Transformer (GPAT), a transformer-based model architecture that accurately predicts part poses by inferring how each part shape corresponds to the target shape. Our experiments on both 3D CAD models and real-world scans demonstrate GPAT's generalization abilities to novel and diverse target and part shapes.

Sumeet Singh, Francis McCann Ramirez, Jacob Varley et al.

Though robot learning is often formulated in terms of discrete-time Markov decision processes (MDPs), physical robots require near-continuous multiscale feedback control. Machines operate on multiple asynchronous sensing modalities, each with different frequencies, e.g., video frames at 30Hz, proprioceptive state at 100Hz, force-torque data at 500Hz, etc. While the classic approach is to batch observations into fixed-time windows then pass them through feed-forward encoders (e.g., with deep networks), we show that there exists a more elegant approach -- one that treats policy learning as modeling latent state dynamics in continuous-time. Specifically, we present 'InFuser', a unified architecture that trains continuous time-policies with Neural Controlled Differential Equations (CDEs). InFuser evolves a single latent state representation over time by (In)tegrating and (Fus)ing multi-sensory observations (arriving at different frequencies), and inferring actions in continuous-time. This enables policies that can react to multi-frequency multi sensory feedback for truly end-to-end visuomotor control, without discrete-time assumptions. Behavior cloning experiments demonstrate that InFuser learns robust policies for dynamic tasks (e.g., swinging a ball into a cup) notably outperforming several baselines in settings where observations from one sensing modality can arrive at much sparser intervals than others.

Soroush Nasiriany, Fei Xia, Wenhao Yu et al.

Vision language models (VLMs) have shown impressive capabilities across a variety of tasks, from logical reasoning to visual understanding. This opens the door to richer interaction with the world, for example robotic control. However, VLMs produce only textual outputs, while robotic control and other spatial tasks require outputting continuous coordinates, actions, or trajectories. How can we enable VLMs to handle such settings without fine-tuning on task-specific data? In this paper, we propose a novel visual prompting approach for VLMs that we call Prompting with Iterative Visual Optimization (PIVOT), which casts tasks as iterative visual question answering. In each iteration, the image is annotated with a visual representation of proposals that the VLM can refer to (e.g., candidate robot actions, localizations, or trajectories). The VLM then selects the best ones for the task. These proposals are iteratively refined, allowing the VLM to eventually zero in on the best available answer. We investigate PIVOT on real-world robotic navigation, real-world manipulation from images, instruction following in simulation, and additional spatial inference tasks such as localization. We find, perhaps surprisingly, that our approach enables zero-shot control of robotic systems without any robot training data, navigation in a variety of environments, and other capabilities. Although current performance is far from perfect, our work highlights potentials and limitations of this new regime and shows a promising approach for Internet-Scale VLMs in robotic and spatial reasoning domains. Website: pivot-prompt.github.io and HuggingFace: https://huggingface.co/spaces/pivot-prompt/pivot-prompt-demo.

Chengshu Li, Jacky Liang, Andy Zeng et al.

Code provides a general syntactic structure to build complex programs and perform precise computations when paired with a code interpreter - we hypothesize that language models (LMs) can leverage code-writing to improve Chain of Thought reasoning not only for logic and arithmetic tasks, but also for semantic ones (and in particular, those that are a mix of both). For example, consider prompting an LM to write code that counts the number of times it detects sarcasm in an essay: the LM may struggle to write an implementation for "detect_sarcasm(string)" that can be executed by the interpreter (handling the edge cases would be insurmountable). However, LMs may still produce a valid solution if they not only write code, but also selectively "emulate" the interpreter by generating the expected output of "detect_sarcasm(string)". In this work, we propose Chain of Code (CoC), a simple yet surprisingly effective extension that improves LM code-driven reasoning. The key idea is to encourage LMs to format semantic sub-tasks in a program as flexible pseudocode that the interpreter can explicitly catch undefined behaviors and hand off to simulate with an LM (as an "LMulator"). Experiments demonstrate that Chain of Code outperforms Chain of Thought and other baselines across a variety of benchmarks; on BIG-Bench Hard, Chain of Code achieves 84%, a gain of 12% over Chain of Thought. In a nutshell, CoC broadens the scope of reasoning questions that LMs can answer by "thinking in code".

Kento Kawaharazuka, Tatsuya Matsushima, Andrew Gambardella et al.

Recent developments in foundation models, like Large Language Models (LLMs) and Vision-Language Models (VLMs), trained on extensive data, facilitate flexible application across different tasks and modalities. Their impact spans various fields, including healthcare, education, and robotics. This paper provides an overview of the practical application of foundation models in real-world robotics, with a primary emphasis on the replacement of specific components within existing robot systems. The summary encompasses the perspective of input-output relationships in foundation models, as well as their role in perception, motion planning, and control within the field of robotics. This paper concludes with a discussion of future challenges and implications for practical robot applications.

Jimmy Wu, Rika Antonova, Adam Kan et al.

For a robot to personalize physical assistance effectively, it must learn user preferences that can be generally reapplied to future scenarios. In this work, we investigate personalization of household cleanup with robots that can tidy up rooms by picking up objects and putting them away. A key challenge is determining the proper place to put each object, as people's preferences can vary greatly depending on personal taste or cultural background. For instance, one person may prefer storing shirts in the drawer, while another may prefer them on the shelf. We aim to build systems that can learn such preferences from just a handful of examples via prior interactions with a particular person. We show that robots can combine language-based planning and perception with the few-shot summarization capabilities of large language models (LLMs) to infer generalized user preferences that are broadly applicable to future interactions. This approach enables fast adaptation and achieves 91.2% accuracy on unseen objects in our benchmark dataset. We also demonstrate our approach on a real-world mobile manipulator called TidyBot, which successfully puts away 85.0% of objects in real-world test scenarios.

Youngsun Wi, Pete Florence, Andy Zeng et al.

Deformable object manipulation requires computationally efficient representations that are compatible with robotic sensing modalities. In this paper, we present VIRDO:an implicit, multi-modal, and continuous representation for deformable-elastic objects. VIRDO operates directly on visual (point cloud) and tactile (reaction forces) modalities and learns rich latent embeddings of contact locations and forces to predict object deformations subject to external contacts.Here, we demonstrate VIRDOs ability to: i) produce high-fidelity cross-modal reconstructions with dense unsupervised correspondences, ii) generalize to unseen contact formations,and iii) state-estimation with partial visio-tactile feedback

Krzysztof Choromanski, Haoxian Chen, Han Lin et al.

We propose a new class of random feature methods for linearizing softmax and Gaussian kernels called hybrid random features (HRFs) that automatically adapt the quality of kernel estimation to provide most accurate approximation in the defined regions of interest. Special instantiations of HRFs lead to well-known methods such as trigonometric (Rahimi and Recht, 2007) or (recently introduced in the context of linear-attention Transformers) positive random features (Choromanski et al., 2021). By generalizing Bochner's Theorem for softmax/Gaussian kernels and leveraging random features for compositional kernels, the HRF-mechanism provides strong theoretical guarantees - unbiased approximation and strictly smaller worst-case relative errors than its counterparts. We conduct exhaustive empirical evaluation of HRF ranging from pointwise kernel estimation experiments, through tests on data admitting clustering structure to benchmarking implicit-attention Transformers (also for downstream Robotics applications), demonstrating its quality in a wide spectrum of machine learning problems.

Michael H. Lim, Andy Zeng, Brian Ichter et al.

Enabling robots to solve multiple manipulation tasks has a wide range of industrial applications. While learning-based approaches enjoy flexibility and generalizability, scaling these approaches to solve such compositional tasks remains a challenge. In this work, we aim to solve multi-task learning through the lens of sequence-conditioning and weighted sampling. First, we propose a new suite of benchmark specifically aimed at compositional tasks, MultiRavens, which allows defining custom task combinations through task modules that are inspired by industrial tasks and exemplify the difficulties in vision-based learning and planning methods. Second, we propose a vision-based end-to-end system architecture, Sequence-Conditioned Transporter Networks, which augments Goal-Conditioned Transporter Networks with sequence-conditioning and weighted sampling and can efficiently learn to solve multi-task long horizon problems. Our analysis suggests that not only the new framework significantly improves pick-and-place performance on novel 10 multi-task benchmark problems, but also the multi-task learning with weighted sampling can vastly improve learning and agent performances on individual tasks.

Pete Florence, Corey Lynch, Andy Zeng et al.

We find that across a wide range of robot policy learning scenarios, treating supervised policy learning with an implicit model generally performs better, on average, than commonly used explicit models. We present extensive experiments on this finding, and we provide both intuitive insight and theoretical arguments distinguishing the properties of implicit models compared to their explicit counterparts, particularly with respect to approximating complex, potentially discontinuous and multi-valued (set-valued) functions. On robotic policy learning tasks we show that implicit behavioral cloning policies with energy-based models (EBM) often outperform common explicit (Mean Square Error, or Mixture Density) behavioral cloning policies, including on tasks with high-dimensional action spaces and visual image inputs. We find these policies provide competitive results or outperform state-of-the-art offline reinforcement learning methods on the challenging human-expert tasks from the D4RL benchmark suite, despite using no reward information. In the real world, robots with implicit policies can learn complex and remarkably subtle behaviors on contact-rich tasks from human demonstrations, including tasks with high combinatorial complexity and tasks requiring 1mm precision.

Lin Yen-Chen, Andy Zeng, Shuran Song et al.

Does having visual priors (e.g. the ability to detect objects) facilitate learning to perform vision-based manipulation (e.g. picking up objects)? We study this problem under the framework of transfer learning, where the model is first trained on a passive vision task, and adapted to perform an active manipulation task. We find that pre-training on vision tasks significantly improves generalization and sample efficiency for learning to manipulate objects. However, realizing these gains requires careful selection of which parts of the model to transfer. Our key insight is that outputs of standard vision models highly correlate with affordance maps commonly used in manipulation. Therefore, we explore directly transferring model parameters from vision networks to affordance prediction networks, and show that this can result in successful zero-shot adaptation, where a robot can pick up certain objects with zero robotic experience. With just a small amount of robotic experience, we can further fine-tune the affordance model to achieve better results. With just 10 minutes of suction experience or 1 hour of grasping experience, our method achieves ~80% success rate at picking up novel objects.

Kevin Zakka, Andy Zeng, Pete Florence et al.

We investigate the visual cross-embodiment imitation setting, in which agents learn policies from videos of other agents (such as humans) demonstrating the same task, but with stark differences in their embodiments -- shape, actions, end-effector dynamics, etc. In this work, we demonstrate that it is possible to automatically discover and learn vision-based reward functions from cross-embodiment demonstration videos that are robust to these differences. Specifically, we present a self-supervised method for Cross-embodiment Inverse Reinforcement Learning (XIRL) that leverages temporal cycle-consistency constraints to learn deep visual embeddings that capture task progression from offline videos of demonstrations across multiple expert agents, each performing the same task differently due to embodiment differences. Prior to our work, producing rewards from self-supervised embeddings typically required alignment with a reference trajectory, which may be difficult to acquire under stark embodiment differences. We show empirically that if the embeddings are aware of task progress, simply taking the negative distance between the current state and goal state in the learned embedding space is useful as a reward for training policies with reinforcement learning. We find our learned reward function not only works for embodiments seen during training, but also generalizes to entirely new embodiments. Additionally, when transferring real-world human demonstrations to a simulated robot, we find that XIRL is more sample efficient than current best methods. Qualitative results, code, and datasets are available at https://x-irl.github.io

Jimmy Wu, Xingyuan Sun, Andy Zeng et al.

The ability to communicate intention enables decentralized multi-agent robots to collaborate while performing physical tasks. In this work, we present spatial intention maps, a new intention representation for multi-agent vision-based deep reinforcement learning that improves coordination between decentralized mobile manipulators. In this representation, each agent's intention is provided to other agents, and rendered into an overhead 2D map aligned with visual observations. This synergizes with the recently proposed spatial action maps framework, in which state and action representations are spatially aligned, providing inductive biases that encourage emergent cooperative behaviors requiring spatial coordination, such as passing objects to each other or avoiding collisions. Experiments across a variety of multi-agent environments, including heterogeneous robot teams with different abilities (lifting, pushing, or throwing), show that incorporating spatial intention maps improves performance for different mobile manipulation tasks while significantly enhancing cooperative behaviors.

Daniel Seita, Pete Florence, Jonathan Tompson et al.

Rearranging and manipulating deformable objects such as cables, fabrics, and bags is a long-standing challenge in robotic manipulation. The complex dynamics and high-dimensional configuration spaces of deformables, compared to rigid objects, make manipulation difficult not only for multi-step planning, but even for goal specification. Goals cannot be as easily specified as rigid object poses, and may involve complex relative spatial relations such as "place the item inside the bag". In this work, we develop a suite of simulated benchmarks with 1D, 2D, and 3D deformable structures, including tasks that involve image-based goal-conditioning and multi-step deformable manipulation. We propose embedding goal-conditioning into Transporter Networks, a recently proposed model architecture for learning robotic manipulation that rearranges deep features to infer displacements that can represent pick and place actions. In simulation and in physical experiments, we demonstrate that goal-conditioned Transporter Networks enable agents to manipulate deformable structures into flexibly specified configurations without test-time visual anchors for target locations. We also significantly extend prior results using Transporter Networks for manipulating deformable objects by testing on tasks with 2D and 3D deformables. Supplementary material is available at https://berkeleyautomation.github.io/bags/.

Andy Zeng, Pete Florence, Jonathan Tompson et al.

Robotic manipulation can be formulated as inducing a sequence of spatial displacements: where the space being moved can encompass an object, part of an object, or end effector. In this work, we propose the Transporter Network, a simple model architecture that rearranges deep features to infer spatial displacements from visual input - which can parameterize robot actions. It makes no assumptions of objectness (e.g. canonical poses, models, or keypoints), it exploits spatial symmetries, and is orders of magnitude more sample efficient than our benchmarked alternatives in learning vision-based manipulation tasks: from stacking a pyramid of blocks, to assembling kits with unseen objects; from manipulating deformable ropes, to pushing piles of small objects with closed-loop feedback. Our method can represent complex multi-modal policy distributions and generalizes to multi-step sequential tasks, as well as 6DoF pick-and-place. Experiments on 10 simulated tasks show that it learns faster and generalizes better than a variety of end-to-end baselines, including policies that use ground-truth object poses. We validate our methods with hardware in the real world. Experiment videos and code are available at https://transporternets.github.io

Jimmy Wu, Xingyuan Sun, Andy Zeng et al.

Typical end-to-end formulations for learning robotic navigation involve predicting a small set of steering command actions (e.g., step forward, turn left, turn right, etc.) from images of the current state (e.g., a bird's-eye view of a SLAM reconstruction). Instead, we show that it can be advantageous to learn with dense action representations defined in the same domain as the state. In this work, we present "spatial action maps," in which the set of possible actions is represented by a pixel map (aligned with the input image of the current state), where each pixel represents a local navigational endpoint at the corresponding scene location. Using ConvNets to infer spatial action maps from state images, action predictions are thereby spatially anchored on local visual features in the scene, enabling significantly faster learning of complex behaviors for mobile manipulation tasks with reinforcement learning. In our experiments, we task a robot with pushing objects to a goal location, and find that policies learned with spatial action maps achieve much better performance than traditional alternatives.

Shuran Song, Andy Zeng, Johnny Lee et al.

Intelligent manipulation benefits from the capacity to flexibly control an end-effector with high degrees of freedom (DoF) and dynamically react to the environment. However, due to the challenges of collecting effective training data and learning efficiently, most grasping algorithms today are limited to top-down movements and open-loop execution. In this work, we propose a new low-cost hardware interface for collecting grasping demonstrations by people in diverse environments. Leveraging this data, we show that it is possible to train a robust end-to-end 6DoF closed-loop grasping model with reinforcement learning that transfers to real robots. A key aspect of our grasping model is that it uses "action-view" based rendering to simulate future states with respect to different possible actions. By evaluating these states using a learned value function (Q-function), our method is able to better select corresponding actions that maximize total rewards (i.e., grasping success). Our final grasping system is able to achieve reliable 6DoF closed-loop grasping of novel objects across various scene configurations, as well as dynamic scenes with moving objects.

Kevin Zakka, Andy Zeng, Johnny Lee et al.

Is it possible to learn policies for robotic assembly that can generalize to new objects? We explore this idea in the context of the kit assembly task. Since classic methods rely heavily on object pose estimation, they often struggle to generalize to new objects without 3D CAD models or task-specific training data. In this work, we propose to formulate the kit assembly task as a shape matching problem, where the goal is to learn a shape descriptor that establishes geometric correspondences between object surfaces and their target placement locations from visual input. This formulation enables the model to acquire a broader understanding of how shapes and surfaces fit together for assembly -- allowing it to generalize to new objects and kits. To obtain training data for our model, we present a self-supervised data-collection pipeline that obtains ground truth object-to-placement correspondences by disassembling complete kits. Our resulting real-world system, Form2Fit, learns effective pick and place strategies for assembling objects into a variety of kits -- achieving $90\%$ average success rates under different initial conditions (e.g. varying object and kit poses), $94\%$ success under new configurations of multiple kits, and over $86\%$ success with completely new objects and kits.

Shreeyak S. Sajjan, Matthew Moore, Mike Pan et al.

Transparent objects are a common part of everyday life, yet they possess unique visual properties that make them incredibly difficult for standard 3D sensors to produce accurate depth estimates for. In many cases, they often appear as noisy or distorted approximations of the surfaces that lie behind them. To address these challenges, we present ClearGrasp -- a deep learning approach for estimating accurate 3D geometry of transparent objects from a single RGB-D image for robotic manipulation. Given a single RGB-D image of transparent objects, ClearGrasp uses deep convolutional networks to infer surface normals, masks of transparent surfaces, and occlusion boundaries. It then uses these outputs to refine the initial depth estimates for all transparent surfaces in the scene. To train and test ClearGrasp, we construct a large-scale synthetic dataset of over 50,000 RGB-D images, as well as a real-world test benchmark with 286 RGB-D images of transparent objects and their ground truth geometries. The experiments demonstrate that ClearGrasp is substantially better than monocular depth estimation baselines and is capable of generalizing to real-world images and novel objects. We also demonstrate that ClearGrasp can be applied out-of-the-box to improve grasping algorithms' performance on transparent objects. Code, data, and benchmarks will be released. Supplementary materials available on the project website: https://sites.google.com/view/cleargrasp

Zhenjia Xu, Jiajun Wu, Andy Zeng et al.

We study the problem of learning physical object representations for robot manipulation. Understanding object physics is critical for successful object manipulation, but also challenging because physical object properties can rarely be inferred from the object's static appearance. In this paper, we propose DensePhysNet, a system that actively executes a sequence of dynamic interactions (e.g., sliding and colliding), and uses a deep predictive model over its visual observations to learn dense, pixel-wise representations that reflect the physical properties of observed objects. Our experiments in both simulation and real settings demonstrate that the learned representations carry rich physical information, and can directly be used to decode physical object properties such as friction and mass. The use of dense representation enables DensePhysNet to generalize well to novel scenes with more objects than in training. With knowledge of object physics, the learned representation also leads to more accurate and efficient manipulation in downstream tasks than the state-of-the-art.

Andy Zeng, Shuran Song, Johnny Lee et al.

We investigate whether a robot arm can learn to pick and throw arbitrary objects into selected boxes quickly and accurately. Throwing has the potential to increase the physical reachability and picking speed of a robot arm. However, precisely throwing arbitrary objects in unstructured settings presents many challenges: from acquiring reliable pre-throw conditions (e.g. initial pose of object in manipulator) to handling varying object-centric properties (e.g. mass distribution, friction, shape) and dynamics (e.g. aerodynamics). In this work, we propose an end-to-end formulation that jointly learns to infer control parameters for grasping and throwing motion primitives from visual observations (images of arbitrary objects in a bin) through trial and error. Within this formulation, we investigate the synergies between grasping and throwing (i.e., learning grasps that enable more accurate throws) and between simulation and deep learning (i.e., using deep networks to predict residuals on top of control parameters predicted by a physics simulator). The resulting system, TossingBot, is able to grasp and throw arbitrary objects into boxes located outside its maximum reach range at 500+ mean picks per hour (600+ grasps per hour with 85% throwing accuracy); and generalizes to new objects and target locations. Videos are available at https://tossingbot.cs.princeton.edu

Andy Zeng, Shuran Song, Stefan Welker et al.

Skilled robotic manipulation benefits from complex synergies between non-prehensile (e.g. pushing) and prehensile (e.g. grasping) actions: pushing can help rearrange cluttered objects to make space for arms and fingers; likewise, grasping can help displace objects to make pushing movements more precise and collision-free. In this work, we demonstrate that it is possible to discover and learn these synergies from scratch through model-free deep reinforcement learning. Our method involves training two fully convolutional networks that map from visual observations to actions: one infers the utility of pushes for a dense pixel-wise sampling of end effector orientations and locations, while the other does the same for grasping. Both networks are trained jointly in a Q-learning framework and are entirely self-supervised by trial and error, where rewards are provided from successful grasps. In this way, our policy learns pushing motions that enable future grasps, while learning grasps that can leverage past pushes. During picking experiments in both simulation and real-world scenarios, we find that our system quickly learns complex behaviors amid challenging cases of clutter, and achieves better grasping success rates and picking efficiencies than baseline alternatives after only a few hours of training. We further demonstrate that our method is capable of generalizing to novel objects. Qualitative results (videos), code, pre-trained models, and simulation environments are available at http://vpg.cs.princeton.edu

Shuran Song, Andy Zeng, Angel X. Chang et al.

We present Im2Pano3D, a convolutional neural network that generates a dense prediction of 3D structure and a probability distribution of semantic labels for a full 360 panoramic view of an indoor scene when given only a partial observation (<= 50%) in the form of an RGB-D image. To make this possible, Im2Pano3D leverages strong contextual priors learned from large-scale synthetic and real-world indoor scenes. To ease the prediction of 3D structure, we propose to parameterize 3D surfaces with their plane equations and train the model to predict these parameters directly. To provide meaningful training supervision, we use multiple loss functions that consider both pixel level accuracy and global context consistency. Experiments demon- strate that Im2Pano3D is able to predict the semantics and 3D structure of the unobserved scene with more than 56% pixel accuracy and less than 0.52m average distance error, which is significantly better than alternative approaches.

Andy Zeng, Shuran Song, Kuan-Ting Yu et al.

This paper presents a robotic pick-and-place system that is capable of grasping and recognizing both known and novel objects in cluttered environments. The key new feature of the system is that it handles a wide range of object categories without needing any task-specific training data for novel objects. To achieve this, it first uses a category-agnostic affordance prediction algorithm to select and execute among four different grasping primitive behaviors. It then recognizes picked objects with a cross-domain image classification framework that matches observed images to product images. Since product images are readily available for a wide range of objects (e.g., from the web), the system works out-of-the-box for novel objects without requiring any additional training data. Exhaustive experimental results demonstrate that our multi-affordance grasping achieves high success rates for a wide variety of objects in clutter, and our recognition algorithm achieves high accuracy for both known and novel grasped objects. The approach was part of the MIT-Princeton Team system that took 1st place in the stowing task at the 2017 Amazon Robotics Challenge. All code, datasets, and pre-trained models are available online at http://arc.cs.princeton.edu

Angel Chang, Angela Dai, Thomas Funkhouser et al.

Access to large, diverse RGB-D datasets is critical for training RGB-D scene understanding algorithms. However, existing datasets still cover only a limited number of views or a restricted scale of spaces. In this paper, we introduce Matterport3D, a large-scale RGB-D dataset containing 10,800 panoramic views from 194,400 RGB-D images of 90 building-scale scenes. Annotations are provided with surface reconstructions, camera poses, and 2D and 3D semantic segmentations. The precise global alignment and comprehensive, diverse panoramic set of views over entire buildings enable a variety of supervised and self-supervised computer vision tasks, including keypoint matching, view overlap prediction, normal prediction from color, semantic segmentation, and region classification.

Shuran Song, Fisher Yu, Andy Zeng et al.

This paper focuses on semantic scene completion, a task for producing a complete 3D voxel representation of volumetric occupancy and semantic labels for a scene from a single-view depth map observation. Previous work has considered scene completion and semantic labeling of depth maps separately. However, we observe that these two problems are tightly intertwined. To leverage the coupled nature of these two tasks, we introduce the semantic scene completion network (SSCNet), an end-to-end 3D convolutional network that takes a single depth image as input and simultaneously outputs occupancy and semantic labels for all voxels in the camera view frustum. Our network uses a dilation-based 3D context module to efficiently expand the receptive field and enable 3D context learning. To train our network, we construct SUNCG - a manually created large-scale dataset of synthetic 3D scenes with dense volumetric annotations. Our experiments demonstrate that the joint model outperforms methods addressing each task in isolation and outperforms alternative approaches on the semantic scene completion task.

Andy Zeng, Kuan-Ting Yu, Shuran Song et al.

Robot warehouse automation has attracted significant interest in recent years, perhaps most visibly in the Amazon Picking Challenge (APC). A fully autonomous warehouse pick-and-place system requires robust vision that reliably recognizes and locates objects amid cluttered environments, self-occlusions, sensor noise, and a large variety of objects. In this paper we present an approach that leverages multi-view RGB-D data and self-supervised, data-driven learning to overcome those difficulties. The approach was part of the MIT-Princeton Team system that took 3rd- and 4th- place in the stowing and picking tasks, respectively at APC 2016. In the proposed approach, we segment and label multiple views of a scene with a fully convolutional neural network, and then fit pre-scanned 3D object models to the resulting segmentation to get the 6D object pose. Training a deep neural network for segmentation typically requires a large amount of training data. We propose a self-supervised method to generate a large labeled dataset without tedious manual segmentation. We demonstrate that our system can reliably estimate the 6D pose of objects under a variety of scenarios. All code, data, and benchmarks are available at http://apc.cs.princeton.edu/

Andy Zeng, Shuran Song, Matthias Nießner et al.

Matching local geometric features on real-world depth images is a challenging task due to the noisy, low-resolution, and incomplete nature of 3D scan data. These difficulties limit the performance of current state-of-art methods, which are typically based on histograms over geometric properties. In this paper, we present 3DMatch, a data-driven model that learns a local volumetric patch descriptor for establishing correspondences between partial 3D data. To amass training data for our model, we propose a self-supervised feature learning method that leverages the millions of correspondence labels found in existing RGB-D reconstructions. Experiments show that our descriptor is not only able to match local geometry in new scenes for reconstruction, but also generalize to different tasks and spatial scales (e.g. instance-level object model alignment for the Amazon Picking Challenge, and mesh surface correspondence). Results show that 3DMatch consistently outperforms other state-of-the-art approaches by a significant margin. Code, data, benchmarks, and pre-trained models are available online at http://3dmatch.cs.princeton.edu