Foundation Models for Rapid Autonomy Validation

We are motivated by the problem of autonomous vehicle performance validation. A key challenge is that an autonomous vehicle requires testing in every kind of driving scenario it could encounter, including rare events, to provide a strong case for safety and show there is no edge-case pathological behavior. Autonomous vehicle companies rely on potentially millions of miles driven in realistic simulation to expose the driving stack to enough miles to estimate rates and severity of collisions. To address scalability and coverage, we propose the use of a behavior foundation model, specifically a masked autoencoder (MAE), trained to reconstruct driving scenarios. We leverage the foundation model in two complementary ways: we (i) use the learned embedding space to group qualitatively similar scenarios together and (ii) fine-tune the model to label scenario difficulty based on the likelihood of a collision upon re-simulation. We use the difficulty scoring as importance weighting for the groups of scenarios. The result is an approach which can more rapidly estimate the rates and severity of collisions by prioritizing hard scenarios while ensuring exposure to every kind of driving scenario.

Online Diffusion-Based 3D Occupancy Prediction at the Frontier with Probabilistic Map Reconciliation

Autonomous navigation and exploration in unmapped environments remains a significant challenge in robotics due to the difficulty robots face in making commonsense inference of unobserved geometries. Recent advancements have demonstrated that generative modeling techniques, particularly diffusion models, can enable systems to infer these geometries from partial observation. In this work, we present implementation details and results for real-time, online occupancy prediction using a modified diffusion model. By removing attention-based visual conditioning and visual feature extraction components, we achieve a 73$\%$ reduction in runtime with minimal accuracy reduction. These modifications enable occupancy prediction across the entire map, rather than being limited to the area around the robot where camera data can be collected. We introduce a probabilistic update method for merging predicted occupancy data into running occupancy maps, resulting in a 71$\%$ improvement in predicting occupancy at map frontiers compared to previous methods. Finally, we release our code and a ROS node for on-robot operation <upon publication> at github.com/arpg/sceneSense_ws.

CogExplore: Contextual Exploration with Language-Encoded Environment Representations

Integrating language models into robotic exploration frameworks improves performance in unmapped environments by providing the ability to reason over semantic groundings, contextual cues, and temporal states. The proposed method employs large language models (GPT-3.5 and Claude Haiku) to reason over these cues and express that reasoning in terms of natural language, which can be used to inform future states. We are motivated by the context of search-and-rescue applications where efficient exploration is critical. We find that by leveraging natural language, semantics, and tracking temporal states, the proposed method greatly reduces exploration path distance and further exposes the need for environment-dependent heuristics. Moreover, the method is highly robust to a variety of environments and noisy vision detections, as shown with a 100% success rate in a series of comprehensive experiments across three different environments conducted in a custom simulation pipeline operating in Unreal Engine.

Radar-Based Localization For Autonomous Ground Vehicles In Suburban Neighborhoods

For autonomous ground vehicles (AGVs) deployed in suburban neighborhoods and other human-centric environments the problem of localization remains a fundamental challenge. There are well established methods for localization with GPS, lidar, and cameras. But even in ideal conditions these have limitations. GPS is not always available and is often not accurate enough on its own, visual methods have difficulty coping with appearance changes due to weather and other factors, and lidar methods are prone to defective solutions due to ambiguous scene geometry. Radar on the other hand is not highly susceptible to these problems, owing in part to its longer range. Further, radar is also robust to challenging conditions that interfere with vision and lidar including fog, smoke, rain, and darkness. We present a radar-based localization system that includes a novel method for highly-accurate radar odometry for smooth, high-frequency relative pose estimation and a novel method for radar-based place recognition and relocalization. We present experiments demonstrating our methods' accuracy and reliability, which are comparable with \new{other methods' published results for radar localization and we find outperform a similar method as ours applied to lidar measurements}. Further, we show our methods are lightweight enough to run on common low-power embedded hardware with ample headroom for other autonomy functions.

SceneSense: Diffusion Models for 3D Occupancy Synthesis from Partial Observation

When exploring new areas, robotic systems generally exclusively plan and execute controls over geometry that has been directly measured. When entering space that was previously obstructed from view such as turning corners in hallways or entering new rooms, robots often pause to plan over the newly observed space. To address this we present SceneScene, a real-time 3D diffusion model for synthesizing 3D occupancy information from partial observations that effectively predicts these occluded or out of view geometries for use in future planning and control frameworks. SceneSense uses a running occupancy map and a single RGB-D camera to generate predicted geometry around the platform at runtime, even when the geometry is occluded or out of view. Our architecture ensures that SceneSense never overwrites observed free or occupied space. By preserving the integrity of the observed map, SceneSense mitigates the risk of corrupting the observed space with generative predictions. While SceneSense is shown to operate well using a single RGB-D camera, the framework is flexible enough to extend to additional modalities. SceneSense operates as part of any system that generates a running occupancy map `out of the box', removing conditioning from the framework. Alternatively, for maximum performance in new modalities, the perception backbone can be replaced and the model retrained for inference in new applications. Unlike existing models that necessitate multiple views and offline scene synthesis, or are focused on filling gaps in observed data, our findings demonstrate that SceneSense is an effective approach to estimating unobserved local occupancy information at runtime. Local occupancy predictions from SceneSense are shown to better represent the ground truth occupancy distribution during the test exploration trajectories than the running occupancy map.

Restorebot: Towards an Autonomous Robotics Platform for Degraded Rangeland Restoration

Degraded rangelands undergo continual shifts in the appearance and distribution of plant life. The nature of these changes however is subtle: between seasons seedlings sprout up and some flourish while others perish, meanwhile, over multiple seasons they experience fluctuating precipitation volumes and can be grazed by livestock. The nature of these conditioning variables makes it difficult for ecologists to quantify the efficacy of intervention techniques under study. To support these observation and intervention tasks, we develop RestoreBot: a mobile robotic platform designed for gathering data in degraded rangelands for the purpose of data collection and intervention in order to support revegetation. Over the course of multiple deployments, we outline the opportunities and challenges of autonomous data collection for revegetation and the importance of further effort in this area. Specifically, we identify that localization, mapping, data association, and terrain assessment remain open problems for deployment, but that recent advances in computer vision, sensing, and autonomy offer promising prospects for autonomous revegetation.

RMap: Millimeter-Wave Radar Mapping Through Volumetric Upsampling

Millimeter Wave Radar is being adopted as a viable alternative to lidar and radar in adverse visually degraded conditions, such as the presence of fog and dust. However, this sensor modality suffers from severe sparsity and noise under nominal conditions, which makes it difficult to use in precise applications such as mapping. This work presents a novel solution to generate accurate 3D maps from sparse radar point clouds. RMap uses a custom generative transformer architecture, UpPoinTr, which upsamples, denoises, and fills the incomplete radar maps to resemble lidar maps. We test this method on the ColoRadar dataset to demonstrate its efficacy.

A Population-Level Analysis of Neural Dynamics in Robust Legged Robots

Recurrent neural network-based reinforcement learning systems are capable of complex motor control tasks such as locomotion and manipulation, however, much of their underlying mechanisms still remain difficult to interpret. Our aim is to leverage computational neuroscience methodologies to understanding the population-level activity of robust robot locomotion controllers. Our investigation begins by analyzing topological structure, discovering that fragile controllers have a higher number of fixed points with unstable directions, resulting in poorer balance when instructed to stand in place. Next, we analyze the forced response of the system by applying targeted neural perturbations along directions of dominant population-level activity. We find evidence that recurrent state dynamics are structured and low-dimensional during walking, which aligns with primate studies. Additionally, when recurrent states are perturbed to zero, fragile agents continue to walk, which is indicative of a stronger reliance on sensory input and weaker recurrence.

Tell Me Where to Go: A Composable Framework for Context-Aware Embodied Robot Navigation

Humans have the remarkable ability to navigate through unfamiliar environments by solely relying on our prior knowledge and descriptions of the environment. For robots to perform the same type of navigation, they need to be able to associate natural language descriptions with their associated physical environment with a limited amount of prior knowledge. Recently, Large Language Models (LLMs) have been able to reason over billions of parameters and utilize them in multi-modal chat-based natural language responses. However, LLMs lack real-world awareness and their outputs are not always predictable. In this work, we develop NavCon, a low-bandwidth framework that solves this lack of real-world generalization by creating an intermediate layer between an LLM and a robot navigation framework in the form of Python code. Our intermediate shoehorns the vast prior knowledge inherent in an LLM model into a series of input and output API instructions that a mobile robot can understand. We evaluate our method across four different environments and command classes on a mobile robot and highlight our NavCon's ability to interpret contextual commands.

Looking Around Corners: Generative Methods in Terrain Extension

In this paper, we provide an early look at our model for generating terrain that is occluded in the initial lidar scan or out of range of the sensor. As a proof of concept, we show that a transformer based framework is able to be overfit to predict the geometries of unobserved roads around intersections or corners. We discuss our method for generating training data, as well as a unique loss function for training our terrain extension network. The framework is tested on data from the SemanticKitti [1] dataset. Unlabeled point clouds measured from an onboard lidar are used as input data to generate predicted road points that are out of range or occluded in the original point-cloud scan. Then the input pointcloud and predicted terrain are concatenated to the terrain-extended pointcloud. We show promising qualitative results from these methods, as well as discussion for potential quantitative metrics to evaluate the overall success of our framework. Finally, we discuss improvements that can be made to the framework for successful generalization to test sets.