CVPR 2022 Papers

We train embodied neural networks to plan and navigate unseen complex 3D environments, emphasising real-world deployment. Rather than requiring prior knowledge of the agent or environment, the planner learns to model the state transitions and rewards. To avoid the potentially hazardous trial-and-error of reinforcement learning, we focus on differentiable planners such as Value Iteration Networks (VIN), which are trained offline from safe expert demonstrations. Although they work well in small simulations, we address two major limitations that hinder their deployment. First, we observed that current differentiable planners struggle to plan long-term in environments with a high branching complexity. While they should ideally learn to assign low rewards to obstacles to avoid collisions, these penalties are not strong enough to guarantee collision-free operation. We thus impose a structural constraint on the value iteration, which explicitly learns to model impossible actions and noisy motion. Secondly, we extend the model to plan exploration with a limited perspective camera under translation and fine rotations, which is crucial for real robot deployment. Our proposals significantly improve semantic navigation and exploration on several 2D and 3D environments, succeeding in settings that are otherwise challenging for differentiable planners. As far as we know, we are the first to successfully apply them to the difficult Active Vision Dataset, consisting of real images captured from a robot.

Following language instructions to navigate in unseen environments is a challenging problem for autonomous embodied agents. The agent not only needs to ground languages in visual scenes, but also should explore the environment to reach its target. In this work, we propose a dual-scale graph transformer (DUET) for joint long-term action planning and fine-grained cross-modal understanding. We build a topological map on-the-fly to enable efficient exploration in global action space. To balance the complexity of large action space reasoning and fine-grained language grounding, we dynamically combine a fine-scale encoding over local observations and a coarse-scale encoding on a global map via graph transformers. The proposed approach, DUET, significantly outperforms state-of-the-art methods on goal-oriented vision-and-language navigation (VLN) benchmarks REVERIE and SOON. It also improves the success rate on the fine-grained VLN benchmark R2R.

Visual exploration is a task that seeks to visit all the navigable areas of an environment as quickly as possible. The existing methods employ deep reinforcement learning (RL) as the standard tool for the task. However, they tend to be vulnerable to statistical shifts between the training and test data, resulting in poor generalization over novel environments that are out-of-distribution (OOD) from the training data. In this paper, we attempt to improve the generalization ability by utilizing the inductive biases available for the task. Employing the active neural SLAM (ANS) that learns exploration policies with the advantage actor-critic (A2C) method as the base framework, we first point out that the mappings represented by the actor and the critic should satisfy specific symmetries. We then propose a network design for the actor and the critic to inherently attain these symmetries. Specifically, we use G-convolution instead of the standard convolution and insert the semi-global polar pooling (SGPP) layer, which we newly design in this study, in the last section of the critic network. Experimental results show that our method increases area coverage by 8.1 square meters when trained on the Gibson dataset and tested on the Matterport3D dataset, establishing the new state-of-the-art.

Contrastive language image pretraining (CLIP) encoders have been shown to be beneficial for a range of visual tasks from classification and detection to captioning and image manipulation. We investigate the effectiveness of CLIP visual backbones for Embodied AI tasks. We build incredibly simple baselines, named EmbCLIP, with no task specific architectures, inductive biases (such as the use of semantic maps), auxiliary tasks during training, or depth maps--yet we find that our improved baselines perform very well across a range of tasks and simulators. EmbCLIP tops the RoboTHOR ObjectNav leaderboard by a huge margin of 20 pts (Success Rate). It tops the iTHOR 1-Phase Rearrangement leaderboard, beating the next best submission, which employs Active Neural Mapping, and more than doubling the % Fixed Strict metric (0.08 to 0.17). It also beats the winners of the 2021 Habitat ObjectNav Challenge, which employ auxiliary tasks, depth maps, and human demonstrations, and those of the 2019 Habitat PointNav Challenge. We evaluate the ability of CLIP's visual representations at capturing semantic information about input observations--primitives that are useful for navigation-heavy embodied tasks--and find that CLIP's representations encode these primitives more effectively than ImageNet-pretrained backbones. Finally, we extend one of our baselines, producing an agent capable of zero-shot object navigation that can navigate to objects that were not used as targets during training. Our code and models are available at https://github.com/allenai/embodied-clip.

Vision-and-language Navigation (VLN) task requires an embodied agent to navigate to a remote location following a natural language instruction. Previous methods usually adopt a sequence model (e.g., Transformer and LSTM) as the navigator. In such a paradigm, the sequence model predicts action at each step through a maintained navigation state, which is generally represented as a one-dimensional vector. However, the crucial navigation clues (i.e., object-level environment layout) for embodied navigation task is discarded since the maintained vector is essentially unstructured. In this paper, we propose a novel Structured state-Evolution (SEvol) model to effectively maintain the environment layout clues for VLN. Specifically, we utilise the graph-based feature to represent the navigation state instead of the vector-based state. Accordingly, we devise a Reinforced Layout clues Miner (RLM) to mine and detect the most crucial layout graph for long-term navigation via a customised reinforcement learning strategy. Moreover, the Structured Evolving Module (SEM) is proposed to maintain the structured graph-based state during navigation, where the state is gradually evolved to learn the object-level spatial-temporal relationship. The experiments on the R2R and R4R datasets show that the proposed SEvol model improves VLN models' performance by large margins, e.g., +3% absolute SPL accuracy for NvEM and +8% for EnvDrop on the R2R test set.

Anticipating future events is an essential feature for intelligent systems and embodied AI. However, compared to the traditional recognition task, the uncertainty of future and reasoning ability requirement make the anticipation task very challenging and far beyond solved. In this filed, previous methods usually care more about the model architecture design or but few attention has been put on how to train an anticipation model with a proper learning policy. To this end, in this work, we propose a novel training scheme called Dynamic Context Removal (DCR), which dynamically schedule the visibility of observed future in the learning procedure. It follows the human-like curriculum learning process, i.e., gradually removing the event context to increase the anticipation difficulty till satisfying the final anticipation target. Our learning scheme is plug-and-play and easy to integrate any reasoning model including transformer and LSTM, with advantages in both effectiveness and efficiency. In extensive experiments, the proposed method achieves state-of-the-art on four widely-used benchmarks. Our code and models are publicly released at https://github.com/AllenXuuu/DCR.

Over the years various methods have been proposed for the problem of object detection. Recently, we have witnessed great strides in this domain owing to the emergence of powerful deep neural networks. However, there are typically two main assumptions common among these approaches. First, the model is trained on a fixed training set and is evaluated on a pre-recorded test set. Second, the model is kept frozen after the training phase, so no further updates are performed after the training is finished. These two assumptions limit the applicability of these methods to real-world settings. In this paper, we propose Interactron, a method for adaptive object detection in an interactive setting, where the goal is to perform object detection in images observed by an embodied agent navigating in different environments. Our idea is to continue training during inference and adapt the model at test time without any explicit supervision via interacting with the environment. Our adaptive object detection model provides a 11.8 point improvement in AP (and 19.1 points in AP50) over DETR, a recent, high-performance object detector. Moreover, we show that our object detection model adapts to environments with completely different appearance characteristics, and its performance is on par with a model trained with full supervision for those environments. We will release the code to help ease future research in this domain.

We present a large-scale study of imitating human demonstrations on tasks that require a virtual robot to search for objects in new environments - (1) ObjectGoal Navigation (e.g. 'find & go to a chair') and (2) Pick&Place (e.g. 'find mug, pick mug, find counter, place mug on counter'). First, we develop a virtual teleoperation data-collection infrastructure - connecting Habitat simulator running in a web browser to Amazon Mechanical Turk, allowing remote users to teleoperate virtual robots, safely and at scale. We collect 80k demonstrations for ObjectNav and 12k demonstrations for Pick&Place, which is an order of magnitude larger than existing human demonstration datasets in simulation or on real robots. Our virtual teleoperation data contains 29.3M actions, and is equivalent to 22.6k hours of real-world teleoperation time, and illustrates rich, diverse strategies for solving the tasks. Second, we use this data to answer the question - how does large-scale imitation learning (IL) (which has not been hitherto possible) compare to reinforcement learning (RL) (which is the status quo)? On ObjectNav, we find that IL (with no bells or whistles) using 70k human demonstrations outperforms RL using 240k agent-gathered trajectories. This effectively establishes an 'exchange rate' - a single human demonstration appears to be worth 4 agent-gathered ones. More importantly, we find the IL-trained agent learns efficient object-search behavior from humans - it peeks into rooms, checks corners for small objects, turns in place to get a panoramic view - none of these are exhibited as prominently by the RL agent, and to induce these behaviors via contemporary RL techniques would require tedious reward engineering. Finally, accuracy vs. training data size plots show promising scaling behavior, suggesting that simply collecting more demonstrations is likely to advance the state of art further. On Pick&Place, the comparison is starker - IL agents achieve 18% success on episodes with new object-receptacle locations when trained with 9.5k human demonstrations, while RL agents fail to get beyond 0%. Overall, our work provides compelling evidence for investing in large-scale imitation learning. Project page: https://ram81.github.io/projects/habitat-web.

3D-aware image generative modeling aims to generate 3D-consistent images with explicitly controllable camera poses. Recent works have shown promising results by training neural radiance field (NeRF) generators on unstructured 2D images, but still cannot generate highly-realistic images with fine details. A critical reason is that the high memory and computation cost of volumetric representation learning greatly restricts the number of point samples for radiance integration during training. Deficient sampling not only limits the expressive power of the generator to handle fine details but also impedes effective GAN training due to the noise caused by unstable Monte Carlo sampling. We propose a novel approach that regulates point sampling and radiance field learning on 2D manifolds, embodied as a set of learned implicit surfaces in the 3D volume. For each viewing ray, we calculate ray-surface intersections and accumulate their radiance generated by the network. By training and rendering such radiance manifolds, our generator can produce high quality images with realistic fine details and strong visual 3D consistency.

Object detection and forecasting are fundamental components of embodied perception. These two problems, however, are largely studied in isolation by the community. In this paper, we propose an end-to-end approach for motion forecasting based on raw sensor measurement as opposed to ground truth tracks. Instead of predicting the current frame locations and forecast forward in time, we directly predict future object locations and backcast in time to determine where each trajectory began. Our approach not only greatly boosts the overall accuracy compared to modular or other end-to-end baselines, it also prompts us to rethink the role of explicit tracking for embodied perception. Additionally, by linking future and current locations in a multiple-to-one manner, our approach is able to reason about multiple futures, a capability that was previously considered difficult for end-to-end approaches. We conduct extensive experiments on the popular autonomous driving dataset nuScenes and demonstrate the empirical effectiveness of our approach. In addition, we investigate the appropriateness of reusing standard forecasting metrics for an end-to-end setup, and find a number of flaws which allow us to build simple baselines to game these metrics. We address this issue with a novel set of joint forecasting and detection metrics that extend the commonly used AP metrics used in the detection community to measuring forecasting accuracy.

The way an object looks and sounds provide complementary reflections of its physical properties. In many settings cues from vision and audition arrive asynchronously but must be integrated, as when we hear an object dropped on the floor and then must find it. In this paper, we introduce a setting in which to study multi-modal object localization in 3D virtual environments. An object is dropped somewhere in a room. An embodied robot agent, equipped with a camera and microphone, must determine what object has been dropped -- and where -- by combining audio and visual signals with knowledge of the underlying physics. To study this problem, we have generated a large-scale dataset -- the Fallen Objects dataset -- that includes 8000 instances of 30 physical object categories in 64 rooms. The dataset uses the ThreeDWorld Platform that can simulate physics-based impact sounds and complex physical interactions between objects in a photorealistic setting. As a first step toward addressing this challenge, we develop a set of embodied agent baselines, based on imitation learning, reinforcement learning, and modular planning, and perform an in-depth analysis of the challenge of this new task.

Digitizing physical objects into the virtual world has the potential to unlock new research and applications in embodied AI and mixed reality. This work focuses on recreating interactive digital twins of real-world articulated objects, which can be directly imported into virtual environments. We introduce Ditto to learn articulation model estimation and 3D geometry reconstruction of an articulated object through interactive perception. Given a pair of visual observations of an articulated object before and after interaction, Ditto reconstructs part-level geometry and estimates the articulation model of the object. We employ implicit neural representations for joint geometry and articulation modeling. Our experiments show that Ditto effectively builds digital twins of articulated objects in a category-agnostic way. We also apply Ditto to real-world objects and deploy the recreated digital twins in physical simulation. Code and additional results are available at https://ut-austin-rpl.github.io/Ditto/

We propose Continuous Scene Representations (CSR), a scene representation constructed by an embodied agent navigating within a space, where objects and their relationships are modeled by continuous valued embeddings. Our method captures feature relationships between objects, composes them into a graph structure on-the-fly, and situates an embodied agent within the representation. Our key insight is to embed pair-wise relationships between objects in a latent space. This allows for a richer representation compared to discrete relations (e.g., [support], [next-to]) commonly used for building scene representations. CSR can track objects as the agent moves in a scene, update the representation accordingly, and detect changes in room configurations. Using CSR, we outperform state-of-the-art approaches for the challenging downstream task of visual room rearrangement, without any task specific training. Moreover, we show the learned embeddings capture salient spatial details of the scene and show applicability to real world data. A summery video and code is available at https://prior.allenai.org/projects/csr.

Vision-Language Navigation (VLN) is a challenging task that requires an embodied agent to perform action-level modality alignment, i.e., make instruction-asked actions sequentially in complex visual environments. Most existing VLN agents learn the instruction-path data directly and cannot sufficiently explore action-level alignment knowledge inside the multi-modal inputs. In this paper, we propose modAlity-aligneD Action PrompTs (ADAPT), which provides the VLN agent with action prompts to enable the explicit learning of action-level modality alignment to pursue successful navigation. Specifically, an action prompt is defined as a modality-aligned pair of an image sub-prompt and a text sub-prompt, where the former is a single-view observation and the latter is a phrase like "walk past the chair". When starting navigation, the instruction-related action prompt set is retrieved from a pre-built action prompt base and passed through a prompt encoder to obtain the prompt feature. Then the prompt feature is concatenated with the original instruction feature and fed to a multi-layer transformer for action prediction. To collect high-quality action prompts into the prompt base, we use the Contrastive Language-Image Pretraining (CLIP) model which has powerful cross-modality alignment ability. A modality alignment loss and a sequential consistency loss are further introduced to enhance the alignment of the action prompt and enforce the agent to focus on the related prompt sequentially. Experimental results on both R2R and RxR show the superiority of ADAPT over state-of-the-art methods.