Simulation can and should play a critical role in the development and testing of algorithms for autonomous agents. What might reduce its impact is the ``sim2real'' gap -- the algorithm response differs between operation in simulated versus real-world environments. This paper introduces an approach to evaluate this gap, focusing on the accuracy of sensor simulation -- specifically IMU and GPS -- in velocity estimation tasks for autonomous agents. Using a scaled autonomous vehicle, we conduct 40 real-world experiments across diverse environments then replicate the experiments in simulation with five distinct sensor noise models. We note that direct comparison of raw simulation and real sensor data fails to quantify the sim2real gap for robotics applications. We demonstrate that by using a state of the art state-estimation package as a ``judge'', and by evaluating the performance of this state-estimator in both real and simulated scenarios, we can isolate the sim2real discrepancies stemming from sensor simulations alone. The dataset generated is open-source and publicly available for unfettered use.
This paper develops a new filtering approach for state estimation in polynomial systems corrupted by arbitrary noise, which commonly arise in robotics. We first consider a batch setup where we perform state estimation using all data collected from the initial to the current time. We formulate the batch state estimation problem as a Polynomial Optimization Problem (POP) and relax the assumption of Gaussian noise by specifying a finite number of moments of the noise. We solve the resulting POP using a moment relaxation and prove that under suitable conditions on the rank of the relaxation, (i) we can extract a provably optimal estimate from the moment relaxation, and (ii) we can obtain a belief representation from the dual (sum-of-squares) relaxation. We then turn our attention to the filtering setup and apply similar insights to develop a GMKF for recursive state estimation in polynomial systems with arbitrary noise. The GMKF formulates the prediction and update steps as POPs and solves them using moment relaxations, carrying over a possibly non-Gaussian belief. In the linear-Gaussian case, GMKF reduces to the standard Kalman Filter. We demonstrate that GMKF performs well under highly non-Gaussian noise and outperforms common alternatives, including the Extended and Unscented Kalman Filter, and their variants on matrix Lie group.
We investigate a variation of the 3D registration problem, named multi-model 3D registration. In the multi-model registration problem, we are given two point clouds picturing a set of objects at different poses (and possibly including points belonging to the background) and we want to simultaneously reconstruct how all objects moved between the two point clouds. This setup generalizes standard 3D registration where one wants to reconstruct a single pose, e.g., the motion of the sensor picturing a static scene. Moreover, it provides a mathematically grounded formulation for relevant robotics applications, e.g., where a depth sensor onboard a robot perceives a dynamic scene and has the goal of estimating its own motion (from the static portion of the scene) while simultaneously recovering the motion of all dynamic objects. We assume a correspondence-based setup where we have putative matches between the two point clouds and consider the practical case where these correspondences are plagued with outliers. We then propose a simple approach based on Expectation-Maximization (EM) and establish theoretical conditions under which the EM approach converges to the ground truth. We evaluate the approach in simulated and real datasets ranging from table-top scenes to self-driving scenarios and demonstrate its effectiveness when combined with state-of-the-art scene flow methods to establish dense correspondences.
Automatic speech recognition (ASR) models are typically trained on large datasets of transcribed speech. As language evolves and new terms come into use, these models can become outdated and stale. In the context of models trained on the server but deployed on edge devices, errors may result from the mismatch between server training data and actual on-device usage. In this work, we seek to continually learn from on-device user corrections through Federated Learning (FL) to address this issue. We explore techniques to target fresh terms that the model has not previously encountered, learn long-tail words, and mitigate catastrophic forgetting. In experimental evaluations, we find that the proposed techniques improve model recognition of fresh terms, while preserving quality on the overall language distribution.
We report on an effort that led to POLAR3D, a set of digital assets that enhance the POLAR dataset of stereo images generated by NASA to mimic lunar lighting conditions. Our contributions are twofold. First, we have annotated each photo in the POLAR dataset, providing approximately 23 000 labels for rocks and their shadows. Second, we digitized several lunar terrain scenarios available in the POLAR dataset. Specifically, by utilizing both the lunar photos and the POLAR's LiDAR point clouds, we constructed detailed obj files for all identifiable assets. POLAR3D is the set of digital assets comprising of rock/shadow labels and obj files associated with the digital twins of lunar terrain scenarios. This new dataset can be used for training perception algorithms for lunar exploration and synthesizing photorealistic images beyond the original POLAR collection. Likewise, the obj assets can be integrated into simulation environments to facilitate realistic rover operations in a digital twin of a POLAR scenario. POLAR3D is publicly available to aid perception algorithm development, camera simulation efforts, and lunar simulation exercises.POLAR3D is publicly available at https://github.com/uwsbel/POLAR-digital.
We report on a study that employs an in-house developed simulation infrastructure to accomplish zero shot policy transferability for a control policy associated with a scale autonomous vehicle. We focus on implementing policies that require no real world data to be trained (Zero-Shot Transfer), and are developed in-house as opposed to being validated by previous works. We do this by implementing a Neural Network (NN) controller that is trained only on a family of circular reference trajectories. The sensors used are RTK-GPS and IMU, the latter for providing heading. The NN controller is trained using either a human driver (via human in the loop simulation), or a Model Predictive Control (MPC) strategy. We demonstrate these two approaches in conjunction with two operation scenarios: the vehicle follows a waypoint-defined trajectory at constant speed; and the vehicle follows a speed profile that changes along the vehicle's waypoint-defined trajectory. The primary contribution of this work is the demonstration of Zero-Shot Transfer in conjunction with a novel feed-forward NN controller trained using a general purpose, in-house developed simulation platform.
Diffusion models have exhibited promising progress in video generation. However, they often struggle to retain consistent details within local regions across frames. One underlying cause is that traditional diffusion models approximate Gaussian noise distribution by utilizing predictive noise, without fully accounting for the impact of inherent information within the input itself. Additionally, these models emphasize the distinction between predictions and references, neglecting information intrinsic to the videos. To address this limitation, inspired by the self-attention mechanism, we propose a novel text-to-video (T2V) generation network structure based on diffusion models, dubbed Additional Perturbation for Latent noise with Adversarial training (APLA). Our approach only necessitates a single video as input and builds upon pre-trained stable diffusion networks. Notably, we introduce an additional compact network, known as the Video Generation Transformer (VGT). This auxiliary component is designed to extract perturbations from the inherent information contained within the input, thereby refining inconsistent pixels during temporal predictions. We leverage a hybrid architecture of transformers and convolutions to compensate for temporal intricacies, enhancing consistency between different frames within the video. Experiments demonstrate a noticeable improvement in the consistency of the generated videos both qualitatively and quantitatively.
Understanding and manipulating articulated objects, such as doors and drawers, is crucial for robots operating in human environments. We wish to develop a system that can learn to articulate novel objects with no prior interaction, after training on other articulated objects. Previous approaches for articulated object manipulation rely on either modular methods which are brittle or end-to-end methods, which lack generalizability. This paper presents FlowBot++, a deep 3D vision-based robotic system that predicts dense per-point motion and dense articulation parameters of articulated objects to assist in downstream manipulation tasks. FlowBot++ introduces a novel per-point representation of the articulated motion and articulation parameters that are combined to produce a more accurate estimate than either method on their own. Simulated experiments on the PartNet-Mobility dataset validate the performance of our system in articulating a wide range of objects, while real-world experiments on real objects' point clouds and a Sawyer robot demonstrate the generalizability and feasibility of our system in real-world scenarios.
Large pre-trained models have had a significant impact on computer vision by enabling multi-modal learning, where the CLIP model has achieved impressive results in image classification, object detection, and semantic segmentation. However, the model's performance on 3D point cloud processing tasks is limited due to the domain gap between depth maps from 3D projection and training images of CLIP. This paper proposes DiffCLIP, a new pre-training framework that incorporates stable diffusion with ControlNet to minimize the domain gap in the visual branch. Additionally, a style-prompt generation module is introduced for few-shot tasks in the textual branch. Extensive experiments on the ModelNet10, ModelNet40, and ScanObjectNN datasets show that DiffCLIP has strong abilities for 3D understanding. By using stable diffusion and style-prompt generation, DiffCLIP achieves an accuracy of 43.2\% for zero-shot classification on OBJ\_BG of ScanObjectNN, which is state-of-the-art performance, and an accuracy of 80.6\% for zero-shot classification on ModelNet10, which is comparable to state-of-the-art performance.
Modeling a robust control system with a precise GPS-based state estimation capability in simulation can be useful in field navigation applications as it allows for testing and validation in a controlled environment. This testing process would enable navigation systems to be developed and optimized in simulation with direct transferability to real-world scenarios. The multi-physics simulation engine Chrono allows for the creation of scenarios that may be difficult or dangerous to replicate in the field, such as extreme weather or terrain conditions. Autonomy Research Testbed (ART), a specialized robotics algorithm testbed, is operated in conjunction with Chrono to develop an MPC control policy as well as an EKF state estimator. This platform enables users to easily integrate custom algorithms in the autonomy stack. This model is initially developed and used in simulation and then tested on a twin vehicle model in reality, to demonstrate the transferability between simulation and reality (also known as Sim2Real).