Abstract: A method learns unsupervised world models for autonomous driving via discrete diffusion. The method includes encoding an observation of an actor for a geographic region using an encoder to generate a prior frame of prior tokens. The method further includes processing the prior frame with a spatio-temporal transformer to generate a predicted frame of predicted tokens. The spatio-temporal transformer includes a spatial transformer and a temporal transformer. The method further includes processing the predicted frame to generate a predicted action for the actor. The method further includes decoding the predicted frame to generate a predicted observation of the geographic region.

Abstract: A method implements multimodal four-dimensional panoptic segmentation. The method includes receiving a set of images and a set of point clouds and executing an image encoder model using the set of images to extract a set of image feature maps. The method further includes executing a point voxel encoder model using the set of image feature maps and the set of point clouds to extract a set of voxel features, a set of image features, and a set of point features and executing a panoptic decoder model using the set of voxel features, the set of image features, the set of point features, and a set of queries to generate a semantic mask and a track mask. The method further includes performing an action responsive to at least one of the semantic mask and the track mask.

Abstract: Latent representation based appearance modification for adversarial testing and training include obtaining a first latent representation of an actor, performing a modification of the first latent representation of an actor to obtain a second latent representation, and generating a 3D model from the second latent representation. The operations further include performing, by a simulator interacting with the virtual driver, a simulation of the virtual world having the 3D model of the actor and the autonomous system moving in the virtual world, evaluating the virtual driver interacting in the virtual world during the simulation to obtain an evaluation result, and outputting the evaluation result.

Abstract: A method implements automatic labeling of objects from LiDAR point clouds via trajectory level refinement. The method includes executing an encoder model using a set of bounding box vectors and a set of point clouds to generate a set of combined feature vectors and executing an attention model using the set of combined feature vectors to generate a set of updated feature vectors. The method further includes executing a decoder model using the set of updated feature vectors to generate a set of pose residuals and a size residual and updating the set of bounding box vectors with the set of pose residuals and the size residual to generate a set of refined bounding box vectors. The method further includes executing an action responsive to the set of refined bounding box vectors.

Abstract: Deferred neural lighting in augmented image generation includes performing operations. The operations include generating a source light representation of a real-world scene from a panoramic image of the real-world scene, augmenting the real-world scene in an object representation of the real-world scene to generate an augmented scene, and processing the augmented scene to generate augmented image buffers. The operations further include selecting a target lighting representation identifying a target light source, processing, by a neural deferred rendering model, the augmented image buffers, the source lighting representation, and a target lighting representation to generate an augmented image having a lighting appearance according to the target light source and outputting the augmented image.

Abstract: Imitation and reinforcement learning for multi-agent simulation includes performing operations. The operations include obtaining a first real-world scenario of agents moving according to first trajectories and simulating the first real-world scenario in a virtual world to generate first simulated states. The simulating includes processing, by an agent model, the first simulated states for the agents to obtain second trajectories. For each of at least a subset of the agents, a difference between a first corresponding trajectory of the agent and a second corresponding trajectory of the agent is calculated and determining an imitation loss is determined based on the difference. The operations further include evaluating the second trajectories according to a reward function to generate a reinforcement learning loss, calculating a total loss as a combination of the imitation loss and the reinforcement learning loss, and updating the agent model using the total loss.

Abstract: A method includes generating a first sample including first raw parameter values of a first modifiable parameters by a probabilistic model and a kernel and executing a first test of a virtual driver of an autonomous system according to the first sample to generate a first evaluation result of multiple evaluation results. The method further includes updating the probabilistic model according to the first evaluation result and training the kernel using the first evaluation result. The method additionally includes generating a second sample including second raw parameter values of the parameters by the probabilistic model and the kernel and executing a second test of a virtual driver of an autonomous system according to the second sample to generate a second evaluation result of the evaluation results. The method further includes presenting the evaluation results.

Abstract: Diffusion for realistic scene generation includes obtaining a current set of agent state vectors and a map data of a geographic region, and iteratively, through multiple diffusion timesteps, updating the current set of agent state vectors. Iteratively updating includes processing, by a noise prediction model, the current set of agent state vectors, a current diffusion timestep of the plurality of diffusion timesteps, and the map data to obtain a noise prediction value, generating a mean using the noise prediction value, generating a distribution function according to the mean, sampling a revised set of agent state vectors from the distribution function, and replacing the current set of agent state vectors with the revised set of agent state vectors. The current set of agent state vectors are outputted.

Abstract: LiDAR based memory segmentation includes obtaining a LiDAR point cloud that includes LiDAR points from a LiDAR sensor, voxelizing the LiDAR points to obtain LiDAR voxels, and encoding the LiDAR voxels to obtain encoded voxels. A LiDAR voxel memory is revised using the encoded voxels to obtain revised LiDAR voxel memory, decoding the revised LiDAR voxel memory to obtain decoded LiDAR voxel memory features. The LiDAR points are segmented using the decoded LiDAR voxel memory features to generate a segmented LiDAR point cloud.

Abstract: Motion planning with implicit occupancy for autonomous systems includes obtaining a set of trajectories through a geographic region for an autonomous system, and generating, for each trajectory in the set of trajectories, a set of points of interest in the geographic region to obtains sets of points of interest. Motion planning further includes quantizing the sets of points of interest to obtain a set of query points in the geographic region and querying the implicit decoder model with the set of query points to obtain point attributes for the set of query points. Motion planning further includes processing, for each trajectory of a least a subset of trajectories, the point attributes corresponding to the set of points of interest to obtain a trajectory cost for the trajectory. From the set of trajectories, a selected trajectory is selected according to trajectory cost.

Abstract: Motion forecasting for autonomous systems includes obtaining map data of a geographic region and historical trajectories of agents located in the geographic region. The map data includes map elements. The agents and the map elements have a corresponding physical locations in the geographic region. Motion forecasting further includes building, from the historical trajectories and the map data, a heterogeneous graph for the agents and the map elements. The heterogeneous graph defines the corresponding physical locations of the agents and the map elements relative to each other of the agents and the map elements. Motion forecasting further includes modelling, by a graph neural network, agent actions of an agent of the agents using the heterogeneous graph to generate an agent goal location, and operating an autonomous system based on the agent goal location.

Abstract: Three dimensional object reconstruction for sensor simulation includes performing operations that include rendering, by a differential rendering engine, an object image from a target object model, and computing, by a loss function of the differential rendering engine, a loss based on a comparison of the object image with an actual image and a comparison of the target object model with a corresponding lidar point cloud. The operations further include updating the target object model by the differential rendering engine according to the loss, and rendering, after updating the target object model, a target object in a virtual world using the target object model.

Abstract: Real world object reconstruction and representation include performing operations that include sampling locations along a camera ray from a virtual camera to a target object to obtain a sample set of the locations along the camera ray. For each location of the at least a subset of the sample set, the operations include determining a position of the location with respect to the target object, executing, based on the position, a reflectance multilayer perceptron (MLP) model, to determine an albedo and material shininess for the location, and computing a radiance for the location and based on a viewing direction of the camera ray using the albedo and the material shininess. The operations further includes rendering a color value for the camera ray by compositing the radiance across the first sample set.

Abstract: Trajectory value learning for autonomous systems includes generating an environment image from sensor input and processing the environment image through an image neural network to obtain a feature map. Trajectory value learning further includes sampling possible trajectories to obtain a candidate trajectory for an autonomous system, extracting, from the feature map, feature vectors corresponding to the candidate trajectory, combining the feature vectors into the input vector, and processing, by a score neural network model, the input vector to obtain a projected score for the candidate trajectory. Trajectory value learning further includes selecting, from the candidate trajectories, the candidate trajectory as a selected trajectory based on the projected score, and implementing the selected trajectory.