Vision-language-action models (VLAs) have shown potential in leveraging pretrained vision-language models and diverse robot demonstrations for learning generalizable sensorimotor control. While this paradigm effectively utilizes large-scale data from both robotic and non-robotic sources, current VLAs primarily focus on direct input–output mappings, lacking the intermediate reasoning steps crucial for complex manipulation tasks. As a result, existing VLAs lack temporal planning or reasoning capabilities. In this paper, we introduce a method that incorporates explicit visual chain-of-thought (CoT) reasoning into vision-language-action models (VLAs) by predicting future image frames autoregressively as visual goals before generating a short action sequence to achieve these goals. We introduce CoT-VLA, a state-of-the-art 7B VLA that can understand and generate visual and action tokens. Our experimental results demonstrate that CoT-VLA achieves strong performance, outperforming the state-of-the-art VLA model by 17% in real-world manipulation tasks and 6% in simulation benchmarks. Videos are available at: https://cot-vla.github.io/.

Vision-based object detectors are a crucial basis for robotics applications as they provide valuable information about object localization in the environment. These need to ensure high reliability in different lighting conditions, occlusions, and visual artifacts, all while running in real-time. Collecting and annotating real-world data for these networks is prohibitively time consuming and costly, especially for custom assets, such as industrial objects, making it untenable for generalization to in-the-wild scenarios. To this end, we present Synthetica, a method for large-scale synthetic data generation for training robust state estimators. This paper focuses on the task of object detection, an important problem which can serve as the front-end for most state estimation problems, such as pose estimation. Leveraging data from a photorealistic ray-tracing renderer, we scale up data generation, generating 2.7 million images, to train highly accurate real-time detection transformers. We present a collection of rendering randomization and training-time data augmentation techniques conducive to robust sim-to-real performance for vision tasks. We demonstrate state-of-the-art performance on the task of object detection while having detectors that run at 50–100Hz which is 9 times faster than the prior state-of-the-art (SOTA). We further demonstrate the usefulness of our training methodology for robotics applications by showcasing a pipeline for use in the real world with custom objects for which there do not exist prior datasets. Our work highlights the importance of scaling synthetic data generation for robust sim-to-real transfer while achieving the fastest real-time inference speeds. Videos and supplementary information can be found at https://sites.google.com/view/synthetica-vision

Robotics policies are always subjected to complex, second order dynamics that entangle their actions with resulting states. In reinforcement learning (RL) contexts, policies have the burden of deciphering these complicated interactions over massive amounts of experience and complex reward functions to learn how to accomplish tasks. Moreover, policies typically issue actions directly to controllers like Operational Space Control (OSC) or joint PD control, which induces straightline motion towards these action targets in task or joint space. However, straightline motion in these spaces for the most part do not capture the rich, nonlinear behavior our robots need to exhibit, shifting the burden of discovering these behaviors more completely to the agent. Unlike these simpler controllers, geometric fabrics capture a much richer and desirable set of behaviors via artificial, second order dynamics grounded in nonlinear geometry. These artificial dynamics shift the uncontrolled dynamics of a robot via an appropriate control law to form behavioral dynamics. Behavioral dynamics unlock a new action space and safe, guiding behavior over which RL policies are trained. Behavioral dynamics enable bang-bang-like RL policy actions that are still safe for real robots, simplify reward engineering, and help sequence real-world, high-performance policies. We describe the framework more generally and create a specific instantiation for the problem of dexterous, in-hand reorientation of a cube by a highly actuated robot hand.

Recent work has demonstrated the ability of deep reinforcement learning (RL) algorithms to learn complex robotic behaviours in simulation, including in the domain of multi-fingered manipulation. However, such models can be challenging to transfer to the real world due to the gap between simulation and reality. In this paper, we present our techniques to train a) a policy that can perform robust dexterous manipulation on an anthropomorphic robot hand and b) a robust pose estimator suitable for providing reliable real-time information on the state of the object being manipulated. Our policies are trained to adapt to a wide range of conditions in simulation. Consequently, our vision-based policies significantly outperform the best vision policies in the literature on the same reorientation task and are competitive with policies that are given privileged state information via motion capture systems. Our work reaffirms the possibilities of sim-to-real transfer for dexterous manipulation in diverse kinds of hardware and simulator setups, and in our case, with the Allegro Hand and Isaac Gym GPU-based simulation. Furthermore, it opens up possibilities for researchers to achieve such results with commonly-available, affordable robot hands and cameras. Videos of the resulting policy and supplementary information, including experiments and demos, can be found on the website.

This paper explores the problem of collision-free motion generation for manipulators by formulating it as a global motion optimization problem. We develop a parallel optimization technique to solve this problem and demonstrate its effectiveness on massively parallel GPUs. We show that combining simple optimization techniques with many parallel seeds leads to solving difficult motion generation problems within 53ms on average, 62x faster than SOTA trajectory optimization methods. We achieve SOTA performance by combining L-BFGS step direction estimation with a novel parallel noisy line search scheme and a particle-based optimization solver. To further aid trajectory optimization, we develop a parallel geometric planner that is atleast 28x faster than SOTA RRTConnect implementations. We also introduce a collision-free IK solver that can solve over 9000 queries/s. We are releasing our GPU accelerated library CuRobo that contains core components for robot motion generation. Additional details are available at sites.google.com/nvidia.com/curobo.

In-hand manipulation of objects is an important capability to enable robots to carry-out tasks which demand high levels of dexterity. This work presents a robot systems approach to learning dexterous manipulation tasks involving moving objects to arbitrary 6-DoF poses. We show empirical benefits, both in simulation and sim - to- real transfer, of using keypoint-based representations for object pose in policy observations and reward calculation to train a model-free reinforcement learning agent. By utilizing domain randomization strategies and large-scale training, we achieve a high success rate of 83 % on a real TriFinger system, with a single policy able to perform grasping, ungrasping, and finger gaiting in order to achieve arbitrary poses within the workspace. We demonstrate that our policy can generalise to unseen objects, and success rates can be further improved through finetuning. With the aim of assisting further research in learning in-hand manipulation, we provide a detailed exposition of our system and make the codebase of our system available, along with checkpoints trained on billions of steps of experience, at https://s2r2-ig.github.io

We introduce DexYCB, a new dataset for capturing hand grasping of objects. We first compare DexYCB with a related one through cross-dataset evaluation. We then present a thorough benchmark of state-of-the-art approaches on three relevant tasks: 2D object and keypoint detection, 6D object pose estimation, and 3D hand pose estimation. Finally, we evaluate a new robotics-relevant task: generating safe robot grasps in human-to-robot object handover.1

To Perform reliably and consistently over sustained periods of time, large-scale automation critically relies on computer simulation. Simulation allows us and supervisory AI to effectively design, validate, and continuously improve complex processes, and helps practitioners to gain insight into the operation and justify future investments. While numerous successful applications of simulation in industry exist, such as circuit simulation, finite element methods, and computeraided design (CAD), state-of-the-art simulators fall short of accurately modeling physical phenomena, such as friction, impact, and deformation.

Teleoperation offers the possibility of imparting robotic systems with sophisticated reasoning skills, intuition, and creativity to perform tasks. However, teleoperation solutions for high degree-of-actuation (DoA), multi-fingered robots are generally cost-prohibitive, while low-cost offerings usually offer reduced degrees of control. Herein, a low-cost, depth-based teleoperation system, DexPilot, was developed that allows for complete control over the full 23 DoA robotic system by merely observing the bare human hand. DexPilot enabled operators to solve a variety of complex manipulation tasks that go beyond simple pick-and-place operations and performance was measured through speed and reliability metrics. DexPilot cost-effectively enables the production of high dimensional, multi-modality, state-action data that can be leveraged in the future to learn sensorimotor policies for challenging manipulation tasks. The videos of the experiments can be found at https://sites.google.com/view/dex-pilot.

Tracking the pose of an object while it is being held and manipulated by a robot hand is difficult for vision-based methods due to significant occlusions. Prior works have explored using contact feedback and particle filters to localize in-hand objects. However, they have mostly focused on the static grasp setting and not when the object is in motion, as doing so requires modeling of complex contact dynamics. In this work, we propose using GPU-accelerated parallel robot simulations and derivative-free, sample-based optimizers to track in-hand object poses with contact feedback during manipulation. We use physics simulation as the forward model for robot-object interactions, and the algorithm jointly optimizes for the state and the parameters of the simulations, so they better match with those of the real world. Our method runs in real-time (30Hz) on a single GPU, and it achieves an average point cloud distance error of 6mm in simulation experiments and 13mm in the real-world ones.

We consider the problem of transferring policies to the real world by training on a distribution of simulated scenarios. Rather than manually tuning the randomization of simulations, we adapt the simulation parameter distribution using a few real world roll-outs interleaved with policy training. In doing so, we are able to change the distribution of simulations to improve the policy transfer by matching the policy behavior in simulation and the real world. We show that policies trained with our method are able to reliably transfer to different robots in two real world tasks: swing-peg-in-hole and opening a cabinet drawer. The video of our experiments can be found at https://sites.google.com/view/simopt.

Current methods for estimating force from tactile sensor signals are either inaccurate analytic models or task-specific learned models. In this paper, we explore learning a robust model that maps tactile sensor signals to force. We specifically explore learning a mapping for the SynTouch BioTac sensor via neural networks. We propose a voxelized input feature layer for spatial signals and leverage information about the sensor surface to regularize the loss function. To learn a robust tactile force model that transfers across tasks, we generate ground truth data from three different sources: (1) the BioTac rigidly mounted to a force torque (FT) sensor, (2) a robot interacting with a ball rigidly attached to the same FT sensor, and (3) through force inference on a planar pushing task by formalizing the mechanics as a system of particles and optimizing over the object motion. A total of 140k samples were collected from the three sources. We achieve a median angular accuracy of 3.5 degrees in predicting force direction (66% improvement over the current state of the art) and a median magnitude accuracy of 0.06 N (93% improvement) on a test dataset. Additionally, we evaluate the learned force model in a force feedback grasp controller performing object lifting and gentle placement. Our results can be found on https: //sites.google.com/view/tactile-force.

To achieve a dexterous robotic manipulation, we need to endow our robot with tactile feedback capability, i.e. the ability to drive action based on tactile sensing. In this paper, we specifically address the challenge of tactile servoing, i.e. given the current tactile sensing and a target/goal tactile sensing - memorized from a successful task execution in the past - what is the action that will bring the current tactile sensing to move closer towards the target tactile sensing at the next time step. We develop a data-driven approach to acquire a dynamics model for tactile servoing by learning from demonstration. Moreover, our method represents the tactile sensing information as to lie on a surface - or a 2D manifold - and perform a manifold learning, making it applicable to any tactile skin geometry. We evaluate our method on a contact point tracking task using a robot equipped with a tactile finger.

Grasping and manipulating objects is an important human skill. Since most objects are designed to be manipulated by human hands, anthropomorphic hands can enable richer human-robot interaction. Desirable grasps are not only stable, but also functional: they enable post-grasp actions with the object. However, functional grasp synthesis for high degree-of-freedom anthropomorphic hands from object shape alone is challenging. We present ContactGrasp, a framework for functional grasp synthesis from object shape and contact on the object surface. Contact can be manually specified or obtained through demonstrations. Our contact representation is object-centric and allows functional grasp synthesis even for hand models different than the one used for demonstration. Using a dataset of contact demonstrations from humans grasping diverse household objects, we synthesize functional grasps for three hand models and two functional intents. The project webpage is https://contactdb.cc.gatech.edu/contactgrasp.html.

Deep learning-based robotic grasping has made significant progress thanks to algorithmic improvements and increased data availability. However, state-of-the-art models are often trained on as few as hundreds or thousands of unique object instances, and as a result generalization can be a challenge. In this work, we explore a novel data generation pipeline for training a deep neural network to perform grasp planning that applies the idea of domain randomization to object synthesis. We generate millions of unique, unrealistic procedurally generated objects, and train a deep neural network to perform grasp planning on these objects. Since the distribution of successful grasps for a given object can be highly multimodal, we propose an autoregressive grasp planning model that maps sensor inputs of a scene to a probability distribution over possible grasps. This model allows us to sample grasps efficiently at test time (or avoid sampling entirely). We evaluate our model architecture and data generation pipeline in simulation and the real world. We find we can achieve a >90% success rate on previously unseen realistic objects at test time in simulation despite having only been trained on random objects. We also demonstrate an 80% success rate on real-world grasp attempts despite having only been trained on random simulated objects.

We introduce SceneNet RGB-D, a dataset providing pixel-perfect ground truth for scene understanding problems such as semantic segmentation, instance segmentation, and object detection. It also provides perfect camera poses and depth data, allowing investigation into geometric computer vision problems such as optical flow, camera pose estimation, and 3D scene labelling tasks. Random sampling permits virtually unlimited scene configurations, and here we provide 5M rendered RGB-D images from 16K randomly generated 3D trajectories in synthetic layouts, with random but physically simulated object configurations. We compare the semantic segmentation performance of network weights produced from pretraining on RGB images from our dataset against generic VGG-16 ImageNet weights. After fine-tuning on the SUN RGB-D and NYUv2 real-world datasets we find in both cases that the synthetically pre-trained network outperforms the VGG-16 weights. When synthetic pre-training includes a depth channel (something ImageNet cannot natively provide) the performance is greater still. This suggests that large-scale high-quality synthetic RGB datasets with task-specific labels can be more useful for pretraining than real-world generic pre-training such as ImageNet. We host the dataset at http://robotvault. bitbucket.io/scenenet-rgbd.html.

Ever more robust, accurate and detailed mapping using visual sensing has proven to be an enabling factor for mobile robots across a wide variety of applications. For the next level of robot intelligence and intuitive user interaction, maps need to extend beyond geometry and appearance - they need to contain semantics. We address this challenge by combining Convolutional Neural Networks (CNNs) and a state-of-the-art dense Simultaneous Localization and Mapping (SLAM) system, ElasticFusion, which provides long-term dense correspondences between frames of indoor RGB-D video even during loopy scanning trajectories. These correspondences allow the CNN's semantic predictions from multiple view points to be probabilistically fused into a map. This not only produces a useful semantic 3D map, but we also show on the NYUv2 dataset that fusing multiple predictions leads to an improvement even in the 2D semantic labelling over baseline single frame predictions. We also show that for a smaller reconstruction dataset with larger variation in prediction viewpoint, the improvement over single frame segmentation increases. Our system is efficient enough to allow real-time interactive use at frame-rates of ≈25Hz.

Most dense RGB/RGB-D SLAM systems require the brightness of 3-D points observed from different viewpoints to be constant. However, in reality, this assumption is difficult to meet even when the surface is Lambertian and illumination is static. One cause is that most cameras automatically tune exposure to adapt to the wide dynamic range of scene radiance, violating the brightness assumption. We describe a novel system - HDRFusion - which turns this apparent drawback into an advantage by fusing LDR frames into an HDR textured volume using a standard RGB-D sensor with auto-exposure (AE) enabled. The key contribution is the use of a normalised metric for frame alignment which is invariant to changes in exposure time. This enables robust tracking in frame-to-model mode and also compensates the exposure accurately so that HDR texture, free of artefacts, can be generated online. We demonstrate that the tracking robustness and accuracy is greatly improved by the approach and that radiance maps can be generated with far greater dynamic range of scene radiance.

Scene understanding is a prerequisite to many high level tasks for any automated intelligent machine operating in real world environments. Recent attempts with supervised learning have shown promise in this direction but also highlighted the need for enormous quantity of supervised data- performance increases in proportion to the amount of data used. However, this quickly becomes prohibitive when considering the manual labour needed to collect such data. In this work, we focus our attention on depth based semantic per-pixel labelling as a scene understanding problem and show the potential of computer graphics to generate virtually unlimited labelled data from synthetic 3D scenes. By carefully synthesizing training data with appropriate noise models we show comparable performance to state-of-the-art RGBD systems on NYUv2 dataset despite using only depth data as input and set a benchmark on depth-based segmentation on SUN RGB-D dataset.

We introduce SceneNet, a framework for generating high-quality annotated 3D scenes to aid indoor scene understanding. SceneNet leverages manually-annotated datasets of real world scenes such as NYUv2 to learn statistics about object co-occurrences and their spatial relationships. Using a hierarchical simulated annealing optimisation, these statistics are exploited to generate a potentially unlimited number of new annotated scenes, by sampling objects from various existing databases of 3D objects such as ModelNet, and textures such as OpenSurfaces and ArchiveTextures. Depending on the task, SceneNet can be used directly in the form of annotated 3D models for supervised training and 3D reconstruction benchmarking, or in the form of rendered annotated sequences of RGB-D frames or videos.

We introduce the Imperial College London and National University of Ireland Maynooth (ICL-NUIM) dataset for the evaluation of visual odometry, 3D reconstruction and SLAM algorithms that typically use RGB-D data. We present a collection of handheld RGB-D camera sequences within synthetically generated environments. RGB-D sequences with perfect ground truth poses are provided as well as a ground truth surface model that enables a method of quantitatively evaluating the final map or surface reconstruction accuracy. Care has been taken to simulate typically observed real-world artefacts in the synthetic imagery by modelling sensor noise in both RGB and depth data. While this dataset is useful for the evaluation of visual odometry and SLAM trajectory estimation, our main focus is on providing a method to benchmark the surface reconstruction accuracy which to date has been missing in the RGB-D community despite the plethora of ground truth RGB-D datasets available.

In matching tasks in computer vision, and particularly in real-time tracking from video, there are generally strong priors available on absolute and relative correspondence locations thanks to motion and scene models. While these priors are often partially used post-hoc to resolve matching consensus in algorithms like RANSAC, it was recently shown that fully integrating them in an `Active Matching' (AM) approach permits efficient guided image processing with rigorous decisions guided by Information Theory. AM's weakness was that the overhead induced by intermediate Bayesian updates required meant poor scaling to cases where many correspondences were sought. In this paper we show that relaxation of the rigid probabilistic model of AM, where every feature measurement directly affects the prediction of every other, permits dramatically more scalable operation without affecting accuracy. We take a general graph-theoretic view of the structure of prior information in matching to sparsify and approximate the interconnections. We demonstrate the performance of two variations, CLAM and SubAM, in the context of sequential camera tracking. These algorithms are highly competitive with other techniques at matching hundreds of features per frame while retaining great intuitive appeal and the full probabilistic capability to digest prior information.