World action models generate actions together with visual predictions of their consequences. Sampling several pairs creates a choice. We study when these predictions provide useful evidence for making that choice.
A network of CCTV cameras and an edge server coordinates four different robots in a working warehouse, with perception, planning, and control handled offboard. The system uses an uncalibrated camera network to navigate narrow aisles, avoid collisions, and manage shared routes without task-specific sensors on the robots.
PriorEye gives end-to-end driving models visual context from street-level images captured along their planned route at other times. A dual-memory module combines these location-specific priors with a learned fallback, improving driving performance and resilience when onboard sensors or retrieved images are unreliable.
This warehouse study investigates how configuration-driven orchestration can simplify the deployment of robots, cameras, and computing infrastructure. Working with Omniscient Innovations and Gompels Healthcare, we compared a shared configuration model with manual setup, finding shorter deployment times and fewer configuration errors as systems grew.
A network of external cameras guides a mobile robot across an office floor without onboard sensing, SLAM, or camera calibration. The system learns connections between camera views by observing people, then hands control between cameras so the robot can navigate through multiple rooms and corridors.
LikePhys evaluates whether video diffusion models distinguish physically plausible events from impossible ones, without additional training. It compares model likelihoods for paired videos across twelve scenarios spanning rigid bodies, fluids, deformable materials, and optics, providing a measure that aligns with human judgements of physical plausibility.
ComBo combines features from different layers of multiple frozen vision foundation models using a lightweight, trainable adapter. It brings complementary representations together for downstream tasks and identifies which models contribute most, improving on individual backbones without fine-tuning the foundation models themselves.
Select2Plan uses pretrained vision-language models for robot navigation without task-specific training or fine-tuning. By combining visual question answering with relevant examples retrieved from a memory bank, it plans actions from either onboard or external camera views and adapts to new environments with a small number of demonstrations.
GraphSCENE generates traffic scenarios for testing autonomous vehicles in simulation. It uses temporal graph neural networks and real-world traffic patterns to create scenarios guided by user requirements, then renders them in CARLA to support systematic testing of challenging interactions between road users.
Visual DNA represents images and datasets through distributions of neuron activations in a frozen neural network. These compact representations allow comparisons that focus on selected visual properties, supporting dataset analysis, evaluation of synthetic imagery, and the selection of data for transfer learning.