Large language models (LLMs) have unlocked new capabilities of task planning from human instructions. However, prior attempts to apply LLMs to real-world robotic tasks are limited by the lack of grounding in the surrounding scene. In this paper, we develop NLMap, an open-vocabulary and queryable scene representation to address this problem. NLMap serves as a framework to gather and integrate contextual information into LLM planners, allowing them to see and query available objects in the scene before generating a context-conditioned plan. NLMap first establishes a natural language queryable scene representation with Visual Language models (VLMs). An LLM based object proposal module parses instructions and proposes involved objects to query the scene representation for object availability and location. An LLM planner then plans with such information about the scene. NLMap allows robots to operate without a fixed list of objects nor executable options, enabling real robot operation unachievable by previous methods. Project website: https://nlmap-saycan.github.io.

This paper introduces Action Image, a new grasp proposal representation that allows learning an end-to-end deep-grasping policy. Our model achieves 84% grasp success on 172 real world objects while being trained only in simulation on 48 objects with just naive domain randomization. Similar to computer vision problems, such as object detection, Action Image builds on the idea that object features are invariant to translation in image space. Therefore, grasp quality is invariant when evaluating the object-gripper relationship; a successful grasp for an object depends on its local context, but is independent of the surrounding environment. Action Image represents a grasp proposal as an image and uses a deep convolutional network to infer grasp quality. We show that by using an Action Image representation, trained networks are able to extract local, salient features of grasping tasks that generalize across different objects and environments. We show that this representation works on a variety of inputs, including color images (RGB), depth images (D), and combined color-depth (RGB-D). Our experimental results demonstrate that networks utilizing an Action Image representation exhibit strong domain transfer between training on simulated data and inference on real-world sensor streams. Finally, our experiments show that a network trained with Action Image improves grasp success (84% vs. 53%) over a baseline model with the same structure, but using actions encoded as vectors.

While robot learning has demonstrated promising results for enabling robots to automatically acquire new skills, a critical challenge in deploying learning-based systems is scale: acquiring enough data for the robot to effectively generalize broadly. Imitation learning, in particular, has remained a stable and powerful approach for robot learning, but critically relies on expert operators for data collection. In this work, we target this challenge, aiming to build an imitation learning system that can continuously improve through autonomous data collection, while simultaneously avoiding the explicit use of reinforcement learning, to maintain the stability, simplicity, and scalability of supervised imitation. To accomplish this, we cast the problem of imitation with autonomous improvement into a multi-task setting. We utilize the insight that, in a multi-task setting, a failed attempt at one task might represent a successful attempt at another task. This allows us to leverage the robot's own trials as demonstrations for tasks other than the one that the robot actually attempted. Using an initial dataset of multitask demonstration data, the robot autonomously collects trials which are only sparsely labeled with a binary indication of whether the trial accomplished any useful task or not. We then embed the trials into a learned latent space of tasks, trained using only the initial demonstration dataset, to draw similarities between various trials, enabling the robot to achieve one-shot generalization to new tasks. In contrast to prior imitation learning approaches, our method can autonomously collect data with sparse supervision for continuous improvement, and in contrast to reinforcement learning algorithms, our method can effectively improve from sparse, task-agnostic reward signals.

Grasping objects under uncertainty remains an open problem in robotics research. This uncertainty is often due to noisy or partial observations of the object pose or shape. To enable a robot to react appropriately to unforeseen effects, it is crucial that it continuously takes sensor feedback into account. While visual feedback is important for inferring a grasp pose and reaching for an object, contact feedback offers valuable information during manipulation and grasp acquisition. In this paper, we use model-free deep reinforcement learning to synthesize control policies that exploit contact sensing to generate robust grasping under uncertainty. We demonstrate our approach on a multi-fingered hand that exhibits more complex finger coordination than the commonly used two-fingered grippers. We conduct extensive experiments in order to assess the performance of the learned policies, with and without contact sensing. While it is possible to learn grasping policies without contact sensing, our results suggest that contact feedback allows for a significant improvement of grasping robustness under object pose uncertainty and for objects with a complex shape.

The promise of learning to learn for robotics rests on the hope that by extracting some information about the learning process itself we can speed up subsequent similar learning tasks. Here, we introduce a computationally efficient online meta-learning algorithm that builds and optimizes a memory model of the optimal learning rate landscape from previously observed gradient behaviors. While performing task specific optimization, this memory of learning rates predicts how to scale currently observed gradients. After applying the gradient scaling our meta-learner updates its internal memory based on the observed effect its prediction had. Our meta-learner can be combined with any gradient-based optimizer, learns on the fly and can be transferred to new optimization tasks. In our evaluations we show that our meta-learning algorithm speeds up learning of MNIST classification and a variety of learning control tasks, either in batch or online learning settings.

Modern robotics is gravitating toward increasingly collaborative human robot interaction. Tools such as acceleration policies can naturally support the realization of reactive, adaptive, and compliant robots. These tools require us to model the system dynamics accurately - a difficult task. The fundamental problem remains that simulation and reality diverge-we do not know how to accurately change a robot's state. Thus, recent research on improving inverse dynamics models has been focused on making use of machine learning techniques. Traditional learning techniques train on the actual realized accelerations, instead of the policy's desired accelerations, which is an indirect data source. Here we show how an additional training signal - measured at the desired accelerations - can be derived from a feedback control signal. This effectively creates a second data source for learning inverse dynamics models. Furthermore, we show how both the traditional and this new data source, can be used to train task-specific models of the inverse dynamics, when used independently or combined. We analyze the use of both data sources in simulation and demonstrate its effectiveness on a real-world robotic platform. We show that our system incrementally improves the learned inverse dynamics model, and when using both data sources combined converges more consistently and faster.

We aim to reliably predict whether a grasp on a known object is successful before it is executed in the real world. There is an entire suite of grasp metrics that has already been developed which rely on precisely known contact points between object and hand. However, it remains unclear whether and how they may be combined into a general purpose grasp stability predictor. In this paper, we analyze these questions by leveraging a large scale database of simulated grasps on a wide variety of objects. For each grasp, we compute the value of seven metrics. Each grasp is annotated by human subjects with ground truth stability labels. Given this data set, we train several classification methods to find out whether there is some underlying, non-trivial structure in the data that is difficult to model manually but can be learned. Quantitative and qualitative results show the complexity of the prediction problem. We found that a good prediction performance critically depends on using a combination of metrics as input features. Furthermore, non-parametric and non-linear classifiers best capture the structure in the data.

We propose a novel method that enables a robot to identify a graspable object part of an unknown object given only noisy and partial information that is obtained from an RGB-D camera. Our method combines the benefits of local with the advantages of global methods. It learns a classifier that takes a local shape representation as input and outputs the probability that a grasp applied at this location will be successful. Given a query data point that is classified in this way, we can retrieve all the locally similar training data points and use them to predict latent global object shape. This information may help to further prune positively labeled grasp hypotheses based on, e.g. relation to the predicted average global shape or suitability for a specific task. This prediction can also guide scene exploration to prune object shape hypotheses. To learn the function that maps local shape to grasp stability we use a Random Forest Classifier. We show that our method reaches the same classification performance as the current state-of-the-art on this dataset which uses a Convolutional Neural Network. Additionally, we exploit the natural ability of the Random Forest to cluster similar data. For a positively predicted query data point, we retrieve all the locally similar training data points that are associated with the same leaf nodes of the Random Forest. The main insight from this work is that local object shape that affords a grasp is also a good predictor of global object shape. We empirically support this claim with quantitative experiments. Additionally, we demonstrate the predictive capability of the method on some real data examples.

In this paper, we consider the problem of robotic grasping of objects when only partial and noisy sensor data of the environment is available. We are specifically interested in the problem of reliably selecting the best hypothesis from a whole set. This is commonly the case when trying to grasp an object for which we can only observe a partial point cloud from one viewpoint through noisy sensors. There will be many possible ways to successfully grasp this object, and even more which will fail. We propose a supervised learning method that is trained with a ranking loss. This explicitly encourages that the top-ranked training grasp in a hypothesis set is also positively labeled. We show how we adapt the standard ranking loss to work with data that has binary labels and explain the benefits of this formulation. Additionally, we show how we can efficiently optimize this loss with stochastic gradient descent. In quantitative experiments, we show that we can outperform previous models by a large margin.

To achieve accurate vision-based control with a robotic arm, a good hand-eye coordination is required. However, knowing the current configuration of the arm can be very difficult due to noisy readings from joint encoders or an inaccurate hand-eye calibration. We propose an approach for robot arm pose estimation that uses depth images of the arm as input to directly estimate angular joint positions. This is a frame-by-frame method which does not rely on good initialisation of the solution from the previous frames or knowledge from the joint encoders. For estimation, we employ a random regression forest which is trained on synthetically generated data. We compare different training objectives of the forest and also analyse the influence of prior segmentation of the arms on accuracy. We show that this approach improves previous work both in terms of computational complexity and accuracy. Despite being trained on synthetic data only, we demonstrate that the estimation also works on real depth images.

One of the central tasks for a household robot is searching for specific objects. It does not only require localizing the target object but also identifying promising search locations in the scene if the target is not immediately visible. As computation time and hardware resources are usually limited in robotics, it is desirable to avoid expensive visual processing steps that are exhaustively applied over the entire image. The human visual system can quickly select those image locations that have to be processed in detail for a given task. This allows us to cope with huge amounts of information and to efficiently deploy the limited capacities of our visual system. In this paper, we therefore propose to use human fixation data to train a top-down saliency model that predicts relevant image locations when searching for specific objects. We show that the learned model can successfully prune bounding box proposals without rejecting the ground truth object locations. In this aspect, the proposed model outperforms a model that is trained only on the ground truth segmentations of the target object instead of fixation data.

Learning of inverse dynamics modeling errors is key for compliant or force control when analytical models are only rough approximations. Thus, designing real time capable function approximation algorithms has been a necessary focus towards the goal of online model learning. However, because these approaches learn a mapping from actual state and acceleration to torque, good tracking is required to observe data points on the desired path. Recently it has been shown how online gradient descent on a simple modeling error offset term to minimize tracking at acceleration level can address this issue. However, to adapt to larger errors a high learning rate of the online learner is required, resulting in reduced compliancy. Thus, here we propose to combine both approaches: The online adapted offset term ensures good tracking such that a nonlinear function approximator is able to learn an error model on the desired trajectory. This, in turn, reduces the load on the adaptive feedback, enabling it to use a lower learning rate. Combined this creates a controller with variable feedback and low gains, and a feedforward model that can account for larger modeling errors. We demonstrate the effectiveness of this framework, in simulation and on a real system.

The caffe framework is one of the leading deep learning toolboxes in the machine learning and computer vision community. While it offers efficiency and configurability, it falls short of a full interface to Python. With increasingly involved procedures for training deep networks and reaching depths of hundreds of layers, creating configuration files and keeping them consistent becomes an error prone process. We introduce the barrista framework, offering full, pythonic control over caffe. It separates responsibilities and offers code to solve frequently occurring tasks for pre-processing, training and model inspection. It is compatible to all caffe versions since mid 2015 and can import and export .prototxt files. Examples are included, e.g., a deep residual network implemented in only 172 lines (for arbitrary depths), comparing to 2320 lines in the official implementation for the equivalent model.

We propose a new large-scale database containing grasps that are applied to a large set of objects from numerous categories. These grasps are generated in simulation and are annotated with different grasp stability metrics. We use a descriptive and efficient representation of the local object shape at which each grasp is applied. Given this data, we present a two-fold analysis: (i) We use crowdsourcing to analyze the correlation of the metrics with grasp success as predicted by humans. The results show that the metric based on physics simulation is a more consistent predictor for grasp success than the standard υ-metric. The results also support the hypothesis that human labels are not required for good ground truth grasp data. Instead the physics-metric can be used to generate datasets in simulation that may then be used to bootstrap learning in the real world. (ii) We apply a deep learning method and show that it can better leverage the large-scale database for prediction of grasp success compared to logistic regression. Furthermore, the results suggest that labels based on the physics-metric are less noisy than those from the υ-metric and therefore lead to a better classification performance.

Parametric filters, such as the Extended Kalman Filter and the Unscented Kalman Filter, typically scale well with the dimensionality of the problem, but they are known to fail if the posterior state distribution cannot be closely approximated by a density of the assumed parametric form.

In this work, we present a new software environment for the comparative evaluation of algorithms for grasping and dexterous manipulation. The key aspect in its development is to provide a tool that allows the reproduction of well-defined experiments in real-life scenarios in every laboratory and, hence, benchmarks that pave the way for objective comparison and competition in the field of grasping. In order to achieve this, experiments are performed on a sound open-source software platform with an extendable structure in order to be able to include a wider range of benchmarks defined by robotics researchers. The environment is integrated into the OpenGRASP toolkit that is built upon the OpenRAVE project and includes grasp-specific extensions and a tool for the creation/integration of new robot models. Currently, benchmarks for grasp and motion planningare included as case studies, as well as a library of domestic everyday objects models, and a real-life scenario that features a humanoid robot acting in a kitchen.