Contrastive vision-language models (VLMs), like CLIP, have gained popularity for their versatile applicability to various downstream tasks. Despite their successes in some tasks, like zero-shot object recognition, they perform surprisingly poor on other tasks, like attribute recognition. Previous work has attributed these challenges to the modality gap, a separation of image and text in the shared representation space, and to a bias towards objects over other factors, such as attributes. In this analysis paper, we investigate both phenomena thoroughly. We evaluated off-the-shelf VLMs and while the gap's influence on performance is typically overshadowed by other factors, we find indications that closing the gap indeed leads to improvements. Moreover, we find that, contrary to intuition, only few embedding dimensions drive the gap and that the embedding spaces are differently organized. To allow for a clean study of object bias, we introduce a definition and a corresponding measure of it. Equipped with this tool, we find that object bias does not lead to worse performance on other concepts, such as attributes per se. However, why do both phenomena, modality gap and object bias, emerge in the first place? To answer this fundamental question and uncover some of the inner workings of contrastive VLMs, we conducted experiments that allowed us to control the amount of shared information between the modalities. These experiments revealed that the driving factor behind both the modality gap and the object bias, is an information imbalance between images and captions, and unveiled an intriguing connection between the modality gap and entropy of the logits.

The remarkable generalization performance of contrastive vision-language models like CLIP is often attributed to the diversity of their training distributions. However, key questions remain unanswered: Can CLIP generalize to an entirely unseen domain when trained on a diverse mixture of domains (domain generalization)? Can it generalize to unseen classes within partially seen domains (compositional generalization)? What factors affect such generalization? To answer these questions, we trained CLIP models on systematically constructed training distributions with controlled domain diversity and object class exposure. Our experiments show that domain diversity is essential for both domain and compositional generalization, yet compositional generalization can be surprisingly weaker than domain generalization when the training distribution contains a suboptimal subset of the test domain. Through data-centric *and* mechanistic analyses, we find that successful generalization requires the learning of sufficiently shared representations in intermediate layers and circuits.

Sample efficiency is a crucial problem in deep reinforcement learning. Recent algorithms, such as REDQ and DroQ, found a way to improve the sample efficiency by increasing the update-to-data (UTD) ratio to 20 gradient update steps on the critic per environment sample. However, this comes at the expense of a greatly increased computational cost. To reduce this computational burden, we introduce CrossQ: A lightweight algorithm for continuous control tasks that makes careful use of Batch Normalization and removes target networks to surpass the current state-of-the-art in sample efficiency while maintaining a low UTD ratio of 1. Notably, CrossQ does not rely on advanced bias-reduction schemes used in current methods. CrossQ's contributions are threefold: (1) it matches or surpasses current state-of-the-art methods in terms of sample efficiency, (2) it substantially reduces the computational cost compared to REDQ and DroQ, (3) it is easy to implement, requiring just a few lines of code on top of SAC.

Learning-based manipulation policies from image inputs often show weak task transfer capabilities. In contrast, visual servoing methods allow efficient task transfer in high-precision scenarios while requiring only a few demonstrations. In this work, we present a framework that formulates the visual servoing task as graph traversal. Our method not only extends the robustness of visual servoing, but also enables multitask capability based on a few task-specific demonstrations. We construct demonstration graphs by splitting existing demonstrations and recombining them. In order to traverse the demonstration graph in the inference case, we utilize a similarity function that helps select the best demonstration for a specific task. This enables us to compute the shortest path through the graph. Ultimately, we show that recombining demonstrations leads to higher task-respective success. We present extensive simulation and real-world experimental results that demonstrate the efficacy of our approach.

Teaching robots new skills quickly and conveniently is crucial for the broader adoption of robotic systems. In this work, we address the problem of one-shot imitation from a single human demonstration, given by an RGB-D video recording. We propose a two-stage process. In the first stage we extract the demonstration trajectory offline. This entails segmenting manipulated objects and determining their relative motion in relation to secondary objects such as containers. In the online trajectory generation stage, we first re-detect all objects, then warp the demonstration trajectory to the current scene and execute it on the robot. To complete these steps, our method leverages several ancillary models, including those for segmentation, relative object pose estimation, and grasp prediction. We systematically evaluate different combinations of correspondence and re-detection methods to validate our design decision across a diverse range of tasks. Specifically, we collect and quantitatively test on demonstrations of ten different tasks including pick-and-place tasks as well as articulated object manipulation. Finally, we perform extensive evaluations on a real robot system to demonstrate the effectiveness and utility of our approach in real-world scenarios. We make the code publicly available at http://ditto.cs.uni-freiburg.de.

Visual Servoing has been effectively used to move a robot into specific target locations or to track a recorded demonstration. It does not require manual programming, but it is typically limited to settings where one demonstration maps to one environment state. We propose a modular approach to extend visual servoing to scenarios with multiple demonstration sequences. We call this conditional servoing, as we choose the next demonstration conditioned on the observation of the robot. This method presents an appealing strategy to tackle multi-step problems, as individual demonstrations can be combined flexibly into a control policy. We propose different selection functions and compare them on a shape-sorting task in simulation. With the reprojection error yielding the best overall results, we implement this selection function on a real robot and show the efficacy of the proposed conditional servoing. For videos of our experiments, please check out our project page: https://lmb.informatik.uni-freiburg.de/projects/conditional_servoing/

Video generation models have become increasingly popular in the last few years, however the standard 2D architectures used today lack natural spatio-temporal modelling capabilities. In this paper, we present a network architecture for video generation that models spatio-temporal consistency without resorting to costly 3D architectures. The architecture facilitates information exchange between neighboring time points, which improves the temporal consistency of both the high level structure as well as the low-level details of the generated frames. The approach achieves state-of-the- art quantitative performance, as measured by the inception score on the UCF-101 dataset as well as better qualitative results. We also introduce a new quantitative measure (S3) that uses downstream tasks for evaluation. Moreover, we present a new multi-label dataset MaisToy, which enables us to evaluate the generalization of the model.

We propose Adaptive Curriculum Generation from Demonstrations (ACGD) for reinforcement learning in the presence of sparse rewards. Rather than designing shaped reward functions, ACGD adaptively sets the appropriate task difficulty for the learner by controlling where to sample from the demonstration trajectories and which set of simulation parameters to use. We show that training vision-based control policies in simulation while gradually increasing the difficulty of the task via ACGD improves the policy transfer to the real world. The degree of domain randomization is also gradually increased through the task difficulty. We demonstrate zero-shot transfer for two real-world manipulation tasks: pick-and-stow and block stacking. A video showing the results can be found at https://lmb.informatik.uni-freiburg.de/projects/curriculum/.

One-shot imitation is the vision of robot programming from a single demonstration, rather than by tedious construction of computer code. We present a practical method for realizing one-shot imitation for manipulation tasks, exploiting modern learning-based optical flow to perform real-time visual servoing. Our approach, which we call FlowControl, continuously tracks a demonstration video, using a specified foreground mask to attend to an object of interest. Using RGB-D observations, FlowControl requires no 3D object models, and is easy to set up. FlowControl inherits great robustness to visual appearance from decades of work in optical flow. We exhibit FlowControl on a range of problems, including ones requiring very precise motions, and ones requiring the ability to generalize.

Estimating 3D hand pose from single RGB images is a highly ambiguous problem that relies on an unbiased training dataset. In this paper, we analyze cross-dataset generalization when training on existing datasets. We find that approaches perform well on the datasets they are trained on, but do not generalize to other datasets or in-the-wild scenarios. As a consequence, we introduce the first large-scale, multi-view hand dataset that is accompanied by both 3D hand pose and shape annotations. For annotating this real-world dataset, we propose an iterative, semi-automated `human-in-the-loop' approach, which includes hand fitting optimization to infer both the 3D pose and shape for each sample. We show that methods trained on our dataset consistently perform well when tested on other datasets. Moreover, the dataset allows us to train a network that predicts the full articulated hand shape from a single RGB image. The evaluation set can serve as a benchmark for articulated hand shape estimation.

This paper presents an approach for computing power grasps for hands with kinematic structure similar to the human hand, which allows the implementation of strategies inspired in human grasping actions. The proposed method first samples the object surface to look for the best spots for creating an opposing grasp with two or three fingers, and then aligns the other fingers to match the local curvature of the object surface. Different grasp strategies are considered, depending on the relative size of the object with respect to the hand, and on the location of potential obstacles in the environment. Several application examples are provided with two different hand models.